Open-source model drops

It predicts labels for new rows in tabular data for classification and regression tasks. The model is available in three sizes from 28M to 215M parameters.

huggingface.cohuggingface-blog29 Sept 2026, 15:30 UTC

It performs mathematical proof verification and image-to-text processing. It runs using the transformers library and safetensors.

huggingface.cohf-models-internlm29 Sept 2026, 06:53 UTC

• 32GB VRAM

Produces speech and audio from text inputs. Requires 32GB VRAM, an AMD Radeon AI PRO R9700, and ROCm 7.2.

huggingface.cohf-models-new-tts28 Sept 2026, 21:50 UTC

The model performs novel view synthesis via image-to-image generation. The model requires the pytorch library and uses safetensors weights.

huggingface.cohf-models-kyutai29 Sept 2026, 09:24 UTC

The benchmark monitors whether a model's performance declines after it is released. It requires a Claude Max subscription and headless Claude Code.

github.comhn-frontpage29 Sept 2026, 22:36 UTC

The model produces ImageNet-1K classification, COCO object detection, and ADE20K semantic segmentation. It requires the Hugging Face CLI to download weights and the code repository for model definitions.

huggingface.cohf-models-trending29 Sept 2026, 07:28 UTC

The model produces text output. No hardware requirements are listed because open weights are not available.

reddit.comreddit-localllama27 Sept 2026, 23:43 UTC

• 32GB VRAM

Produces text-to-speech audio and voice. Requires 32GB VRAM, an AMD Radeon AI PRO R9700, and ROCm 7.2.

huggingface.cohf-models-new-tts28 Sept 2026, 21:50 UTC

• Apache-2.0

It produces zero-shot voice-cloning and text-to-speech audio for Badini Kurdish. It requires a short reference audio clip and an exact transcript for each generation.

huggingface.cohf-models-new-tts26 Sept 2026, 20:26 UTC

• 8-BIT

The model performs text generation, vision-language processing, audio, and video understanding. The model is available in 8-bit and fp8 formats.

huggingface.cohf-models-xiaomi27 Sept 2026, 04:01 UTC

• 8-BIT

The model generates text from multimodal inputs including audio, vision, and video. The model is available in 8-bit.

huggingface.cohf-models-xiaomi27 Sept 2026, 03:58 UTC

• MIT
• CC-BY-NC
• non-commercial licence

The model performs audio-to-audio voice conversion for Chinese content. Running requires the model weight files and source code.

huggingface.cohf-models-new-audio-to-audio26 Sept 2026, 03:31 UTC

The model produces text and performs on nonogram logic puzzles. The model is an open-weight release.

reddit.comreddit-localllama27 Sept 2026, 10:51 UTC

• 1B params

The model produces text. Available versions include 1B and 27B parameters.

reddit.comreddit-localllama27 Sept 2026, 08:36 UTC

• Apache-2.0

The model produces zero-shot voice-cloned audio for Badini Kurdish. It requires a reference clip and a transcript for each generation.

huggingface.cohf-models-new-tts26 Sept 2026, 20:26 UTC

• Apache-2.0

It produces diagrams from a text language where the user specifies the placement of elements. The software is released under an Apache-2.0 license.

github.comhn-frontpage26 Sept 2026, 17:10 UTC

The Bonsai LLM provides vision and language processing. The 2-billion-parameter model runs on the Snapdragon AR1 Gen 1 Platform.

techcrunch.comtechcrunch-ai24 Sept 2026, 19:00 UTC

Produces audio from Japanese text input. The model uses XTTS architecture and is licensed under agpl-3.0.

huggingface.cohf-models-new-tts24 Sept 2026, 14:33 UTC

• 16B total / 3B active
• 16B params

The repository provides training-content summaries and documentation for model versions. The model features 16B total parameters and 3B active parameters.

huggingface.cohf-models-inclusionai24 Sept 2026, 08:37 UTC

• 3B params
• GGUF

Produces vision-language output by generating a block of candidate tokens from shared image and text representations. The model has 3B parameters and supports GGUF, llama.cpp, MLX-VLM, and SGLang.

huggingface.cohuggingface-blog24 Sept 2026, 14:08 UTC

The 2-billion-parameter model processes vision and language tasks for smart glasses. The model runs on Snapdragon AR1 Gen 1 Platform hardware.

techcrunch.comtechcrunch-ai24 Sept 2026, 19:00 UTC

• 4B params
• Apache-2.0

It generates text based on a Qwen3.5-4B fine-tune. The model has 4B parameters and an Apache-2.0 license.

reddit.comreddit-localllama25 Sept 2026, 17:30 UTC

• 16B total / 3B active
• 16B params
• Apache-2.0

The model performs joint image understanding and generation. The model contains 16B total parameters and 3B active parameters.

huggingface.cohf-models-new-multimodal24 Sept 2026, 07:56 UTC

Produces Japanese text-to-speech audio. Requires an XTTS inference environment.

huggingface.cohf-models-new-tts24 Sept 2026, 14:33 UTC

• 16B total / 3B active
• 16B params

The repository provides training-content summaries and documentation. The repository contains documentation, not model weights.

huggingface.cohf-models-inclusionai24 Sept 2026, 08:37 UTC

• Apache-2.0

It produces decision model outputs in a single forward pass without token-by-token generation. It runs on a local server via an HTTP API.

ollaya.devhn-frontpage25 Sept 2026, 18:33 UTC

The application provides a shared canvas and SDK for humans and agents to architect software. The application runs on macOS and Fedora and integrates with Claude Code, Codex, GPT-6 Sol, and Claude Opus 5.5.

github.comhn-frontpage24 Sept 2026, 17:21 UTC

The tokenizer encodes text and estimates token counts for English, Korean, and Japanese. Requires the Transformers library.

huggingface.cohf-models-upstage21 Sept 2026, 03:08 UTC

• 60 fps

The model generates images and performs tasks via custom agent skills. The system runs at 60 FPS on Ubuntu 24.04+ for ARM64.

github.comunsloth-releases22 Sept 2026, 16:36 UTC

These models produce text and code. These models run on local hardware.

jyn.devhn-frontpage23 Sept 2026, 09:21 UTC

• 30B total / 3B active
• 30B params
• FP8

The model processes video, audio, image, and text for any-to-any tasks. The model contains 31B total parameters, 3B active parameters, and a 256k token context window.

huggingface.cohf-models-new-multimodal21 Sept 2026, 06:17 UTC

• 22.05 kHz

Produces multilingual text-to-speech audio with voice cloning and emotion control. Requires 7.72 GiB of storage and supports 22.05 kHz audio.

huggingface.cohf-models-new-tts22 Sept 2026, 12:03 UTC

It produces text-to-video, video-to-video, image-text-to-video, and video-inpainting using Canny, Depth, HED, MLSD, and Pose control conditions. It runs using the videox_fun library.

huggingface.cohf-models-new-video22 Sept 2026, 03:26 UTC

Produces image-to-video content using transformer weights, camera adapter, and LoRA modules. Requires the WorldCrafter_Fast directory for shared components including the text encoder, VAE, and scheduler.

huggingface.cohf-models-new-i2v21 Sept 2026, 14:30 UTC

Generates image-to-video and text-to-video content with camera control. Requires the WorldCrafter inference code and a pinned uv environment.

huggingface.cohf-models-new-i2v21 Sept 2026, 14:30 UTC

• 7B params

The model produces action chunks and denoised video frames from camera frames, robot states, and text instructions. The model contains 7B parameters.

huggingface.cohf-models-bfl22 Sept 2026, 10:36 UTC

• Apache-2.0

The framework provides tools and building blocks for running local AI. The system requires Apple Silicon hardware and uses the Apache-2.0 license.

huggingface.cohuggingface-blog22 Sept 2026, 00:00 UTC

• 27B params
• GGUF

The Transformers library enables local inference of quantized models. It uses the GGUF format and supports 27B parameter models.

huggingface.cohuggingface-blog22 Sept 2026, 00:00 UTC

This tokenizer encodes text and estimates token counts for English, Korean, and Japanese. Running this requires the Transformers library.

huggingface.cohf-models-upstage21 Sept 2026, 03:08 UTC

• 7B params

Produces image edits and rewritten prompts. 7B parameters.

huggingface.cohf-models-qwen20 Sept 2026, 08:46 UTC

The model performs any-to-any multimodal tasks with long-context and agentic capabilities. No technical specifications are provided.

huggingface.cohf-models-new-multimodal22 Sept 2026, 00:24 UTC

The model produces text and scored 28.8% on Terminal-Bench 4.0. The provided text contains no specific hardware requirements.

reddit.comreddit-localllama21 Sept 2026, 20:54 UTC

• 30B total / 3B active
• 30B params
• FP8

The model performs any-to-any multimodal tasks for video, audio, image, and text. The model uses 31B total parameters and 3B active parameters for a 256k context window.

huggingface.cohf-models-new-multimodal21 Sept 2026, 06:17 UTC

• 0.6B params
• Apache-2.0
• 44.1 kHz
• INT4

Produces multilingual text-to-speech and zero-shot voice cloning. Runs as a 0.6B parameter model with INT4 quantization.

huggingface.cohf-models-new-tts21 Sept 2026, 07:23 UTC

Produces image-to-video and text-to-video content. Requires the diffusers library and a pinned uv environment.

huggingface.cohf-models-new-i2v21 Sept 2026, 14:30 UTC

• Apache-2.0

MLX is a framework for running local AI. The software is optimized for Apple Silicon hardware.

huggingface.cohuggingface-blog22 Sept 2026, 00:00 UTC

• 8-BIT

The model generates text from multimodal inputs including vision, audio, and video. The model runs using 8-bit or fp8 precision.

huggingface.cohf-models-xiaomi21 Sept 2026, 15:39 UTC

• 7B params

Generates images from text prompts. Uses 7B parameters.

huggingface.cohf-models-qwen20 Sept 2026, 08:45 UTC

• Apache-2.0

The model performs any-to-any tasks including conversational text generation and image-to-text processing. The model runs with the transformers library under the Apache-2.0 license.

huggingface.cohf-models-new-multimodal19 Sept 2026, 17:18 UTC

Encodes text and estimates token counts for English, Korean, and Japanese. Requires the Transformers library and contains a vocabulary of 196,608 tokens.

huggingface.cohf-models-upstage21 Sept 2026, 03:08 UTC

• 7B params

This model produces image-to-image transformations and rewritten prompts. The model has 7B parameters.

huggingface.cohf-models-qwen20 Sept 2026, 08:46 UTC

The model produces text. The source text provides no hardware requirements or model size specifications.

reddit.comreddit-localllama21 Sept 2026, 20:54 UTC

• 29B params
• Apache-2.0

The model performs any-to-any tasks including text generation in multiple languages. The model has 29B parameters and utilizes the transformers library.

huggingface.cohf-models-new-multimodal21 Sept 2026, 15:54 UTC

• 0.8B params

Produces yes/no, multiple-choice, and rating outputs for decision-making tasks. Available in 0.8B, 4B, and 9B sizes for CUDA, ROCm, and Apple Silicon hardware.

github.comhn-frontpage21 Sept 2026, 07:11 UTC

• 30B total / 3B active
• 30B params
• FP8

The model generates text, video, audio, and image content. The model uses 31 billion total parameters and 3 billion active parameters per token.

huggingface.cohf-models-new-multimodal21 Sept 2026, 06:17 UTC

• 0.5B params

The model produces text-to-speech audio as an observer head for the CosyVoice2-0.5B native-token interface. The model requires the PyTorch library and has 2,892,653 parameters.

huggingface.cohf-models-new-tts20 Sept 2026, 06:02 UTC

• 0.6B params
• Apache-2.0
• 44.1 kHz
• INT4

This model converts text into multilingual audio and performs zero-shot voice cloning. The model features 0.6B parameters and supports 44.1 kHz audio in INT4 format.

huggingface.cohf-models-new-tts21 Sept 2026, 07:23 UTC

• 8-BIT

Produces text for multimodal, vision-language, audio, and video-understanding tasks. Runs using 8-bit or fp8 precision with the transformers library.

huggingface.cohf-models-xiaomi21 Sept 2026, 15:39 UTC

• 8-BIT

The model performs text generation and processes multimodal inputs including vision, audio, and video. The model supports 8-bit and fp8 quantization and runs with the transformers library.

huggingface.cohf-models-xiaomi21 Sept 2026, 15:39 UTC

• 7B params

This model generates images from text prompts. The model has 7B parameters.

huggingface.cohf-models-qwen20 Sept 2026, 08:45 UTC

• Apache-2.0

This model generates images from text input. It runs via the transformers library under an Apache-2.0 license.

huggingface.cohf-models-new-t2i21 Sept 2026, 14:05 UTC

• Apache-2.0

AIVORENCE published Mia-v1-E2B to the Hugging Face hub for any-to-any. Specs: Apache-2.0.

huggingface.cohf-models-new-multimodal19 Sept 2026, 17:18 UTC

• Apache-2.0

convaiinnovations published laya-multilingual to the Hugging Face hub for text-classification. Specs: Apache-2.0.

huggingface.cohf-models-trending19 Sept 2026, 03:23 UTC

• 7B params

This model performs image editing and prompt rewriting. The model contains 7B parameters.

huggingface.cohf-models-qwen20 Sept 2026, 08:46 UTC

• 26B total / 4B active
• 26B params
• 128K context
• Apache-2.0
• 140 languages

Produces text for any-to-any tasks in 140 languages with a 128K context window. Requires 26B total parameters with 4B active parameters.

huggingface.cohf-models-new-multimodal20 Sept 2026, 17:54 UTC

• Apache-2.0

Pirate Face mirrors Hugging Face models as peer-to-peer torrents with SHA-256 verification. Apache-2.0.

pirateface.cohn-frontpage20 Sept 2026, 15:16 UTC

• 0.5B params

This model produces text-to-speech audio using an observer head for the CosyVoice2-0.5B system. The system requires PyTorch and uses 2,892,653 parameters.

huggingface.cohf-models-new-tts20 Sept 2026, 06:02 UTC

• 7B params
• 166 languages
• INT4

Produces text across 166 languages for any-to-any tasks. Requires 7B parameters and INT4 quantization.

huggingface.cohf-models-new-multimodal18 Sept 2026, 17:43 UTC

• 0.8B params
• 22.05 kHz
• 6GB VRAM

The model produces zero-shot text-to-speech and voice cloning in Chinese, English, Japanese, Spanish, and Arabic with emotion control. The model has 0.8B parameters and requires 6GB VRAM to run at a 22.05 kHz sample rate.

huggingface.cohf-models-new-tts19 Sept 2026, 16:23 UTC

• 7B params
• 6s clips
• 24 kHz
• 17 languages

Produces text-to-speech audio and clones voices into 17 languages from 6-second clips. The model contains 7B parameters and supports 24 kHz audio.

huggingface.cohf-models-new-tts19 Sept 2026, 01:23 UTC

• Apache-2.0
• 15s clips

It produces multi-shot scenes as one continuous take by chaining 10-15 second video blocks with consistent audio and color. The system requires the minimax-h3 library and a ComfyUI node pack.

huggingface.cohf-models-new-video19 Sept 2026, 20:29 UTC

• 7B params

The model generates 2048x2048 images with alpha channels, rendered text, and processes up to 10 input images for editing. The model runs on 7B parameters using an MMDiT architecture on consumer cards.

blog.comfy.orgcomfyui-blog20 Sept 2026, 13:12 UTC

• 7B params

This model generates images from text prompts. The model has 7B parameters.

huggingface.cohf-models-qwen20 Sept 2026, 08:45 UTC

• Apache-2.0

Produces any-to-any outputs including conversational text and image-to-text. Requires the transformers library and safetensors format under an Apache-2.0 license.

huggingface.cohf-models-new-multimodal19 Sept 2026, 17:18 UTC

• Apache-2.0

The model performs text classification for routing, scoring, and moderation. It runs using the transformers library under the Apache-2.0 license.

huggingface.cohf-models-trending18 Sept 2026, 05:05 UTC

• 8B params
• Apache-2.0
• 100 languages

The model produces probability predictions over structured schemas using a non-autoregressive architecture and reinforcement learning. The model has 8B parameters, an Apache-2.0 license, and supports 100 languages.

laya.convaiinnovations.comhn-frontpage19 Sept 2026, 10:46 UTC

• 7B params
• 166 languages
• INT4

The model performs any-to-any text generation in 166 languages. The model has 7B parameters and uses INT4 quantization.

huggingface.cohf-models-new-multimodal18 Sept 2026, 17:43 UTC

• 0.8B params
• 22.05 kHz
• 6GB VRAM

It produces zero-shot text-to-speech with voice cloning and emotion control in Chinese, English, Japanese, Spanish, and Arabic. The model requires 0.8B parameters and 6GB VRAM to produce 22.05 kHz audio.

huggingface.cohf-models-new-tts19 Sept 2026, 16:23 UTC

• 7B params
• 6s clips
• 24 kHz
• 17 languages

Produces text-to-speech and voice cloning in 17 languages from 6-second audio samples. Uses 7B parameters to output 24 kHz audio.

huggingface.cohf-models-new-tts19 Sept 2026, 01:23 UTC

This model converts graphemes to phonemes for text-to-speech. It is a .tflite file with fixed-length [1, 96] and FP32 precision for CPU.

huggingface.cohf-models-new-tts17 Sept 2026, 15:04 UTC

The model generates video from image inputs. No hardware specifications are provided.

huggingface.cohf-models-new-i2v16 Sept 2026, 22:47 UTC

The update adds thinking controls to the show API and adds support for Nemotron H vision models on Apple Silicon. Nemotron H vision models run on Apple Silicon via MLX.

github.comollama-releases19 Sept 2026, 08:10 UTC

• Apache-2.0

The model performs text classification. The model requires the transformers library and uses the safetensors format.

huggingface.cohf-models-trending18 Sept 2026, 05:05 UTC

• 3B total / 0.5B active
• 4B params
• Apache-2.0

It functions as a large language model. The model has 4B parameters and is licensed under Apache-2.0.

huggingface.cohf-models-tencent16 Sept 2026, 03:07 UTC

• 27B params
• 262K context
• Apache-2.0

This multimodal model performs reasoning, coding, vision, and agentic tasks. The 27B parameter model requires 5.9GB of memory and supports a 262K context window.

prismml.comhn-frontpage17 Sept 2026, 21:13 UTC

This model converts graphemes to phonemes for the Kokoro text-to-speech system. The model runs as a TFLite file in FP32 format on CPU.

huggingface.cohf-models-new-tts17 Sept 2026, 15:04 UTC

• 7B params
• 6s clips
• 24 kHz
• 17 languages

Produces cloned voices in 17 languages from 6-second audio clips. The model has 7 billion parameters and operates at a 24 kHz sample rate.

huggingface.cohf-models-new-tts16 Sept 2026, 00:22 UTC

This library manages memory allocation, provides bug fixes, and reports pinned-memory usage statistics. No hardware or software requirements are listed in the source text.

github.comhn-frontpage18 Sept 2026, 04:20 UTC

It indexes visited web pages and local files for search via a web interface, terminal, or AI assistant. Requires a downloaded binary, a running terminal server, and a browser extension for Chrome or Firefox.

github.comhn-frontpage17 Sept 2026, 16:25 UTC

• 29B total / 4B active
• 29B params
• Apache-2.0

The model produces text and handles conversational tasks. The model contains 29B total parameters and 4B active parameters.

huggingface.cohf-models-trending16 Sept 2026, 06:44 UTC

Converts graphemes to phonemes for text-to-speech systems. Runs on CPU using the LiteRT framework in FP32 format.

huggingface.cohf-models-new-tts17 Sept 2026, 15:04 UTC

• 7B params
• 6s clips
• 24 kHz
• 17 languages

Produces speech in 17 languages and clones voices from 6-second audio samples. Contains 7B parameters and outputs 24 kHz audio.

huggingface.cohf-models-new-tts16 Sept 2026, 00:22 UTC

• 5B params
• 24 fps

It generates video sequences based on keyboard inputs and text instructions. The model has 5B parameters and produces output at 24 fps.

arxiv.orgarxiv-cv17 Sept 2026, 04:00 UTC

• CC-BY-NC

Provides 1,271 sentence-aligned pairs translated from Wolof into Modern Standard Arabic for machine translation. Available under a CC-BY-NC license.

arxiv.orgarxiv-cl17 Sept 2026, 04:00 UTC

It indexes visited web pages and local files for retrieval via a web interface, terminal, or AI assistant connected through MCP. It requires a local binary, a running listener process, and a browser extension for Firefox or Chrome.

github.comhn-frontpage17 Sept 2026, 16:25 UTC

• 744B params
• FP8

The model generates text in English and Chinese. The model has 744B parameters and uses FP8 precision.

huggingface.cohf-models-internlm15 Sept 2026, 12:47 UTC

• 7B params
• 6s clips
• 24 kHz
• 17 languages

Produces speech in 17 languages and clones voices from 6-second audio clips. The model has 7B parameters and 24 kHz audio output.

huggingface.cohf-models-new-tts16 Sept 2026, 00:22 UTC

• Apache-2.0
• 24 kHz

Produces Turkish speech from text using a speaker-agnostic acoustic model. Requires the antalia library and supports 24 kHz audio.

huggingface.cohf-models-new-tts14 Sept 2026, 09:57 UTC

The software produces images and text through a node-based interface. The system supports a 4TB AMD Windows VA quota.

github.comcomfyui-releases15 Sept 2026, 22:26 UTC

• 744B params
• FP8

Produces text in Chinese and English. Requires hardware to run 744B parameters in FP8.

huggingface.cohf-models-internlm12 Sept 2026, 04:01 UTC

Produces emotionally expressive and duration-controlled zero-shot text-to-speech in English and Chinese. The model uses the safetensors format.

huggingface.cohf-models-new-tts12 Sept 2026, 09:57 UTC

• 770B params
• FP8

This release provides inference for Hy4-preview, Tencent 770B/49B-active MoE, Qwen3.8-Flash-Next, GraniteSWA, GraniteMoeSWA, and NemotronH_Omni_Reasoning_V3. It supports 770B and 49B parameter models using FP8, BF16, and NVFP4.

github.comvllm-releases10 Sept 2026, 23:53 UTC

• 780B params

Produces text output with a 1M token context window. The model has 780B total parameters and activates 49B parameters per token.

github.comtransformers-releases10 Sept 2026, 12:03 UTC

• 30B total / 3B active
• 176.94B params

Fine-tunes multiple model families using a declarative system, ternary QAT, and PyTorch 2.13. Qwen3.8-Flash-Next 176.94B MoE requires 120 GiB on a single B300.

github.comaxolotl-releases10 Sept 2026, 14:19 UTC

• 1.53B params
• Apache-2.0
• non-commercial licence
• 24 kHz

Produces text-to-speech audio, speech-editing, and audio-to-audio content. Requires a 1.53B parameter DiT, 24 kHz sampling rate, and 6.76 GB of storage.

huggingface.cohf-models-new-tts10 Sept 2026, 14:42 UTC

• 5.84B params
• 2M context
• Apache-2.0

The model produces outputs for any-to-any tasks including OCR, speech-recognition, text-to-speech, and video. The model has 5.84 billion parameters and a 2 million token context window.

huggingface.cohf-models-new-multimodal9 Sept 2026, 13:46 UTC

• 0.1B params
• 44.1 kHz
• INT8

Produces text-to-speech audio and voice samples. Requires 0.1B parameters with 44.1 kHz sampling and INT8 quantization.

huggingface.cohf-models-new-tts11 Sept 2026, 02:20 UTC

• 0.6B params
• Apache-2.0
• 44.1 kHz
• INT4

It produces multilingual text-to-speech and zero-shot voice cloning. The model has 0.6B parameters and supports 44.1 kHz audio in INT4 format.

huggingface.cohf-models-new-tts10 Sept 2026, 15:05 UTC

• 3B params
• CC-BY-NC

Generates music and voice audio from text input. 3B parameters with a CC-BY-NC-4.0 license.

huggingface.cohf-models-map9 Sept 2026, 18:30 UTC

• 3B params
• CC-BY-NC
• 48 kHz

The model performs feature extraction for audio. The model contains 3B parameters and operates at 48 kHz.

huggingface.cohf-models-map9 Sept 2026, 18:30 UTC

• 3B params
• CC-BY-NC
• 30s clips
• 24 kHz

The model produces music representations from 30s audio clips. The model has 632M parameters and 24 layers for 24 kHz mono audio.

huggingface.cohf-models-map9 Sept 2026, 18:28 UTC

The app provides a messaging interface with Material 3 components and large-screen support. No hardware requirements are provided.

github.comhn-frontpage11 Sept 2026, 18:50 UTC

• Apache-2.0

It produces text output from image and text inputs for conversational and reasoning tasks. It runs using the transformers library and supports safetensors format under an Apache-2.0 license.

huggingface.cohf-models-trending11 Sept 2026, 15:50 UTC

• 770B params
• FP8

vLLM provides inference for text models including MoE, Qwen, and Nemotron architectures. The software supports FP8, BF16, and NVFP4 precision on CUDA and ROCm hardware.

github.comvllm-releases10 Sept 2026, 23:53 UTC

• 780B params

The model generates text output. The model has 780B parameters, 49B active parameters per token, and a 1M token context window.

github.comtransformers-releases10 Sept 2026, 12:03 UTC

• 30B total / 3B active
• 176.94B params

It produces fine-tuned language models and multimodal models. It requires a single B300 with 120 GiB of memory to fine-tune a 176.94B parameter model.

github.comaxolotl-releases10 Sept 2026, 14:19 UTC

• 1.53B params
• Apache-2.0
• non-commercial licence
• 24 kHz

This model produces text-to-speech audio. The 1.53B parameter model requires the maestro library.

huggingface.cohf-models-new-tts10 Sept 2026, 14:42 UTC

• 5.84B params
• 2M context
• Apache-2.0

The model produces text, speech-recognition, and text-to-speech outputs for any-to-any tasks. The model has 5.84 billion parameters and a 2 million token context window.

huggingface.cohf-models-new-multimodal9 Sept 2026, 13:46 UTC

• 0.1B params
• 44.1 kHz
• INT8

Produces multilingual text-to-speech audio and voice from text input. The model has 0.1B parameters and supports 44.1 kHz audio in INT8 format.

huggingface.cohf-models-new-tts11 Sept 2026, 02:20 UTC

• 0.6B params
• Apache-2.0
• 44.1 kHz
• INT4

The model produces multilingual text-to-speech audio and zero-shot voice cloning. The model has 0.6B parameters and supports INT4 quantization.

huggingface.cohf-models-new-tts10 Sept 2026, 15:05 UTC

• 3B params
• CC-BY-NC

The model generates music and audio from text inputs. The model has 3 billion parameters.

huggingface.cohf-models-map9 Sept 2026, 18:30 UTC

• 3B params
• CC-BY-NC
• 48 kHz

This model performs feature extraction and functions as an audio autoencoder. The model contains 3B parameters and supports 48 kHz audio.

huggingface.cohf-models-map9 Sept 2026, 18:30 UTC

• 3B params
• CC-BY-NC
• 30s clips
• 24 kHz

The model produces feature representations from 30-second audio clips sampled at 24 kHz. The model contains 632 million parameters across 24 layers and 1,024 dimensions.

huggingface.cohf-models-map9 Sept 2026, 18:28 UTC

• 780B params

The model produces text with a 1M token context window. The model requires infrastructure for 780B total parameters and 49B active parameters per token.

github.comtransformers-releases10 Sept 2026, 12:03 UTC

• 552B params
• 8-BIT

The model produces text from image and text inputs. The model has 552B parameters and supports 8-bit precision.

huggingface.cohf-models-deepseek10 Sept 2026, 02:17 UTC

• 770B params
• FP8

This update implements Model Runner V2 and adds support for models including Hy4-preview, Qwen3.8-Flash-Next, and NemotronH_Omni_Reasoning_V3. It supports 770B and 49B parameter models using FP8, BF16, and NVFP4 formats.

github.comvllm-releases9 Sept 2026, 09:40 UTC

• 3B params

Mistral today announced that it has raised €3 billion in a Series D funding round at a post-money valuation of more than €21 billion. Specs: 3B params.

mistral.aimistral8 Sept 2026, 12:00 UTC

• 124B params
• FP8

The model produces text output from image and text inputs. The model has 124B parameters and uses FP8 precision.

huggingface.cohf-models-inclusionai8 Sept 2026, 10:21 UTC

• 5.84B params
• 2M context
• Apache-2.0

The model performs any-to-any tasks including OCR, speech-recognition, text-to-speech, and video. It features 5.84 billion parameters and a 2 million token context window.

huggingface.cohf-models-new-multimodal9 Sept 2026, 13:46 UTC

• 3B params
• CC-BY-NC

It produces music and audio from text. The model has 3 billion parameters.

huggingface.cohf-models-map9 Sept 2026, 18:30 UTC

• 3B params
• CC-BY-NC
• 48 kHz

This model performs feature extraction for audio data using an autoencoder. The model has 3B parameters and supports 48 kHz audio.

huggingface.cohf-models-map9 Sept 2026, 18:30 UTC

• 3B params
• CC-BY-NC
• 30s clips
• 24 kHz

Extracts 1,024-dimension music representations from 30-second audio clips. Requires 632 million parameters and 24 kHz mono audio input.

huggingface.cohf-models-map9 Sept 2026, 18:28 UTC

It produces web interfaces. It is a CSS framework.

tailwindcss.comhn-frontpage9 Sept 2026, 13:27 UTC

• 780B params

The model produces text outputs with a 1M token context window. The model contains 780B parameters and activates 49B parameters per token.

github.comtransformers-releases9 Sept 2026, 15:42 UTC

• Apache-2.0

The model produces zero-shot time-series forecasts and performs missing value imputation. The model is a 385M-parameter model released under an Apache 2.0 license.

huggingface.cohuggingface-blog9 Sept 2026, 15:36 UTC

• 770B params
• FP8

It serves text generation for models using Model Runner V2 and batch-sharded sampling. Supports 770B parameters in FP8 and 49B active MoE configurations.

github.comvllm-releases9 Sept 2026, 09:40 UTC

• 3B params

This large language model produces text. The model has 3 billion parameters.

mistral.aimistral8 Sept 2026, 12:00 UTC

• 5.84B params
• 2M context
• Apache-2.0

The model produces any-to-any outputs including OCR, speech recognition, text-to-speech, and video. The model has 5.84 billion parameters and a 2 million token context window.

huggingface.cohf-models-new-multimodal9 Sept 2026, 13:46 UTC

• Apache-2.0
• non-commercial licence

The model produces text-to-speech audio in English and Chinese. It runs using PyTorch, CUDA, and the transformers library.

huggingface.cohf-models-new-tts9 Sept 2026, 12:23 UTC

This model produces text-to-speech audio for Rioplatense Spanish using zero-shot flow matching and a Vocos vocoder. The model has 19.8M parameters and was trained on 8 hours of Argentinian Spanish.

huggingface.cohf-models-new-tts9 Sept 2026, 06:24 UTC

• Apache-2.0
• 16 kHz

The model converts text into Eastern Armenian speech. It utilizes a 512-dimensional speaker embedding and 16 kHz audio.

huggingface.cohf-models-new-tts8 Sept 2026, 18:41 UTC

Ollama models can be used in ChatGPT Desktop with support for tool search and response compaction. Setup is available on MacOS and the release includes improvements for Apple Silicon.

github.comollama-releases9 Sept 2026, 17:57 UTC

Tailwind produces CSS styles for web interface development. The text contains no information regarding hardware or system requirements.

tailwindcss.comhn-frontpage9 Sept 2026, 13:27 UTC

• 3B params

Mistral produces text outputs. The model has 3 billion parameters.

mistral.aimistral8 Sept 2026, 12:00 UTC

Produces text generation, math, and long-chain-of-thought reasoning. Requires the transformers and pytorch libraries.

huggingface.cohf-models-openbmb7 Sept 2026, 08:25 UTC

• 3s clips

The model generates text and executes actions without standard safety safeguards. The model is an open-weight model that is fast and cheap to run.

jyn.devhn-frontpage8 Sept 2026, 04:48 UTC

• 22.05 kHz

The model converts Vietnamese text into spoken audio. The model produces 22.05 kHz audio using the Matcha-TTS and Vocos frameworks.

huggingface.cohf-models-new-tts7 Sept 2026, 13:27 UTC

Produces voice and audio from text input. Requires ExecuTorch and a react-native-executorch environment.

huggingface.cohf-models-new-tts7 Sept 2026, 06:59 UTC

• CC-BY-NC
• 22.05 kHz

Produces Finnish text-to-speech audio. Requires NumPy and Raspberry Pi hardware at 22.05 kHz.

huggingface.cohf-models-new-tts7 Sept 2026, 05:58 UTC

Ollama models run in ChatGPT Desktop with structured output on Apple Silicon. Requires the Ollama app on MacOS.

github.comollama-releases5 Sept 2026, 23:59 UTC

• 2B params
• Apache-2.0

The model generates text and supports long-context and tool-calling. The model has 2B parameters and an Apache-2.0 license.

huggingface.cohf-models-openbmb6 Sept 2026, 11:33 UTC

The system estimates calorie counts from meal photos and descriptions. The system requires an LLM and the Nutrition5k dataset.

reddit.comreddit-localllama7 Sept 2026, 05:56 UTC

• 26B total / 4B active
• 12B params
• 128K context
• Apache-2.0
• 140 languages

The model produces any-to-any and image-text-to-text content in 140 languages. The model uses 12B total parameters, 4B active parameters, and a 128K context window.

huggingface.cohf-models-new-multimodal4 Sept 2026, 23:00 UTC

• CC-BY-NC
• 22.05 kHz

The model produces Finnish text-to-speech audio. It runs on NumPy and supports a 22.05 kHz sample rate.

huggingface.cohf-models-new-tts7 Sept 2026, 05:58 UTC

• Apache-2.0
• non-commercial licence

The model generates speech and audio from text in English and Chinese. The model runs using the transformers library, pytorch, and CUDA.

huggingface.cohf-models-new-tts6 Sept 2026, 07:12 UTC

• 22.05 kHz

It converts English text into speech using a FastSpeech2 and HiFi-GAN architecture. The model requires PyTorch and has 59.5 million parameters for 22.05 kHz audio.

huggingface.cohf-models-new-tts5 Sept 2026, 17:23 UTC

• CC-BY-NC
• 5s clips
• 24 kHz

The model generates Burmese speech from text input. It requires the f5-tts library and produces 24 kHz audio.

huggingface.cohf-models-new-tts5 Sept 2026, 14:02 UTC

• 11s clips
• 24 kHz
• 4-BIT

Produces Tamil text-to-speech audio for a single female voice. The 86 MB model uses 4-bit quantised weights to run on a CPU without a GPU.

huggingface.cohf-models-new-tts5 Sept 2026, 05:22 UTC

It integrates Ollama models into ChatGPT Desktop and improves structured output performance on Apple Silicon. Requires the Ollama app on MacOS.

github.comollama-releases5 Sept 2026, 23:59 UTC

• Apache-2.0

This model produces text and supports tool calling, multi-turn interaction, and error correction. This model requires the transformers library and utilizes the LLaDA2.0-mini architecture.

huggingface.cohf-models-inclusionai5 Sept 2026, 05:16 UTC

Produces synchronized audio and video in a single generation pass. No specific hardware or software requirements are listed.

blog.comfy.orgcomfyui-blog3 Sept 2026, 22:28 UTC

Integrates Ollama models with ChatGPT Desktop and improves structured output on Apple Silicon. Requires the Ollama app on MacOS.

github.comollama-releases5 Sept 2026, 23:59 UTC

• Apache-2.0
• 4K

It produces dense and late-interaction embeddings from text tokens and raw image patches for multimodal and multilingual retrieval. Available in 260M and 800M sizes, it runs on an NVIDIA L40S GPU with 2048x2048 image inputs.

huggingface.cohuggingface-blog3 Sept 2026, 13:13 UTC

• 26B total / 4B active
• 12B params
• 128K context
• Apache-2.0
• 140 languages

The model produces any-to-any and image-text-to-text results. The model uses 26B total parameters with 4B active and a 128K context window.

huggingface.cohf-models-new-multimodal4 Sept 2026, 09:55 UTC

This model produces text-to-speech audio in English, Japanese, Chinese, and Russian. No hardware or software requirements are listed.

huggingface.cohf-models-new-tts5 Sept 2026, 19:57 UTC

• 22.05 kHz

The system produces English speech from text using FastSpeech2 and HiFi-GAN at 22.05 kHz. The model requires PyTorch and contains 59.5 million parameters.

huggingface.cohf-models-new-tts5 Sept 2026, 17:23 UTC

• CC-BY-NC
• 24 kHz

This model converts text into Burmese speech. It requires the f5-tts library and outputs 24 kHz audio.

huggingface.cohf-models-new-tts5 Sept 2026, 14:02 UTC

• Apache-2.0

The model produces multilingual text-to-speech with streaming output and reference-based voice cloning. The model requires the iceAudio library and an active acoustic adapter.

huggingface.cohf-models-new-tts3 Sept 2026, 14:03 UTC

MiniMax H3 produces synchronized audio and video in a single generation pass. The system runs in ComfyUI.

blog.comfy.orgcomfyui-blog3 Sept 2026, 22:28 UTC

• Apache-2.0
• 4K

We introduce NeoMME, a family of 260M and 800M multilingual multimodal encoders. Specs: Apache-2.0 · 4K.

huggingface.cohuggingface-blog3 Sept 2026, 13:13 UTC

• 26B total / 4B active
• 12B params
• 128K context
• Apache-2.0
• 140 languages

OsGo published gemma-4-E2B to the Hugging Face hub for any-to-any. Specs: 26B total / 4B active · 12B params · 128K context · Apache-2.0 · 140 languages.

huggingface.cohf-models-new-multimodal4 Sept 2026, 09:55 UTC

• 300B params
• Apache-2.0
• GGUF

I'm one of the developers. Specs: 300B params · Apache-2.0 · GGUF.

reddit.comreddit-localllama4 Sept 2026, 09:15 UTC

• non-commercial licence

TimesFM-3 is the third generation of Google Research's zero-shot forecasting model, and the main change from 2.5 is that it handles multivariate inputs natively instead of being limited to a single series' own history. Specs: non-commercial licence.

reddit.comreddit-localllama3 Sept 2026, 19:34 UTC

• 27B params
• 8-BIT

Available Models Model Compression This section provides transparency about the compression state of each model available on our platform. Specs: 27B params · 8-BIT.

inference-docs.cerebras.aihn-frontpage3 Sept 2026, 18:32 UTC

• 5s clips
• 24 kHz

Thanushttk published pocket-tts-hindi to the Hugging Face hub for text-to-speech (voice and audio). Specs: 5s clips · 24 kHz.

huggingface.cohf-models-new-tts4 Sept 2026, 09:59 UTC

jjiang4 published thorsten_vits to the Hugging Face hub for text-to-speech (voice and audio). No specs stated in the source — open the link before quoting numbers.

huggingface.cohf-models-new-tts3 Sept 2026, 02:42 UTC

• GGUF

Run Qwen3.8-Flash-Next and GLM-5.3-Flash up to 2x faster with MTP. Specs: GGUF.

github.comunsloth-releases2 Sept 2026, 13:49 UTC

Two weeks, one rule, and hundreds of entries from nearly 50 countries. No specs stated in the source — open the link before quoting numbers.

blog.comfy.orgcomfyui-blog3 Sept 2026, 22:28 UTC

• Apache-2.0
• 4K

The model produces dense and late-interaction embeddings from text and image patches for visual document retrieval. The model is available in 260M and 800M sizes under the Apache-2.0 license.

huggingface.cohuggingface-blog3 Sept 2026, 13:13 UTC

• non-commercial licence

The model produces multivariate time series forecasts. The model has 330M parameters and a non-commercial license.

reddit.comreddit-localllama3 Sept 2026, 19:34 UTC

• 27B params
• 8-BIT

Produces text output using the Qwen 3.8 27B model. The model has 27B parameters and runs at 1500 tokens per second with 8-bit weights.

inference-docs.cerebras.aihn-frontpage3 Sept 2026, 18:32 UTC

• 375B total / 23B active
• 375B params
• Apache-2.0

The model fleet produces text for reasoning, mathematics, coding, and agentic tasks. The suite includes six models with parameter counts ranging from 0.9B to 375B.

ifm.aihn-frontpage3 Sept 2026, 15:36 UTC

• 36B total / 4B active
• 36B params
• Apache-2.0

The model performs text generation and conversational tasks. The model has 36 billion total parameters and 4 billion active parameters.

huggingface.cohf-models-trending1 Sept 2026, 23:01 UTC

• 26B total / 4B active
• 12B params
• 128K context
• Apache-2.0
• 140 languages

The model produces any-to-any outputs in 140 languages. It has 12B parameters and a 128K context window.

huggingface.cohf-models-new-multimodal1 Sept 2026, 17:12 UTC

The model converts text into Hausa speech audio. It requires the coqui-tts library and XTTS-v2 architecture.

huggingface.cohf-models-new-tts2 Sept 2026, 12:17 UTC

• 7B params
• MIT
• 10 languages

The model performs automatic speech recognition for audio input in 10 languages. The system uses 7B parameters and runs with the transformers library.

huggingface.cohf-models-microsoft2 Sept 2026, 15:46 UTC

• GGUF

The model generates audio output for Qwen and GLM models. The model runs in GGUF format.

github.comunsloth-releases2 Sept 2026, 13:49 UTC

• Apache-2.0

Provides a library of WebGPU kernels and an in-browser benchmarking suite. Licensed under Apache-2.0.

huggingface.cohuggingface-blog1 Sept 2026, 00:00 UTC

Audacity 4.0 provides a clip-editing model, multi-clip selection, grouping, and a split tool. No hardware requirements are provided.

github.comhn-frontpage3 Sept 2026, 10:53 UTC

• 8B params

It produces a question-answer dataset based on NCERT textbooks for Indian education. The model requires 8B parameters.

arxiv.orgarxiv-cl1 Sept 2026, 04:00 UTC

The tool evaluates large language models on benchmark suites. The system runs as a Python-based evaluation harness with ONNX support.

github.comlm-eval-harness-releases31 Aug 2026, 20:42 UTC

• 770B params
• 1M context

The model produces text from text inputs without vision capabilities. The model requires 770B total parameters, 49B active parameters, a 1M context window, and 1.56TB storage.

simonwillison.netsimonwillison29 Aug 2026, 23:53 UTC

• 26B total / 4B active
• 12B params
• 128K context
• Apache-2.0
• 140 languages

Produces text output from text and image inputs across 140 languages. Requires the transformers library to run with 12B parameters and a 128K context window.

huggingface.cohf-models-new-multimodal1 Sept 2026, 17:12 UTC

• 0.1B params
• 44.1 kHz

The model produces zero-shot text-to-speech audio in multiple languages including English, Chinese, German, Spanish, French, Italian, Japanese, and Korean. The model has 0.1B parameters and supports 44.1 kHz audio.

huggingface.cohf-models-new-tts31 Aug 2026, 18:07 UTC

• 0.8B params
• 22.05 kHz
• 6GB VRAM

Produces zero-shot text-to-speech with voice cloning and emotion control in multiple languages. Requires 0.8B parameters and 6GB VRAM to output 22.05 kHz audio.

huggingface.cohf-models-new-tts31 Aug 2026, 15:21 UTC

Produces Dutch text-to-speech audio from input text. Requires the transformers library, the microsoft/speecht5_hifigan vocoder, and a speaker embedding.

huggingface.cohf-models-new-tts31 Aug 2026, 12:19 UTC

• 2.2B params
• Apache-2.0

It converts text into spoken audio. The model has 2.2B parameters and uses the transformers library.

huggingface.cohf-models-new-tts31 Aug 2026, 10:40 UTC

• non-commercial licence

Produces 3D assets from a single image including baked normal and ambient occlusion maps. Runs on consumer hardware in ComfyUI without custom nodes or non-commercial dependencies.

blog.comfy.orgcomfyui-blog31 Aug 2026, 22:01 UTC

• Apache-2.0

Provides supplemental model checkpoints for research purposes. Model weights are available under the Apache-2.0 license.

huggingface.cohf-models-allenai1 Sept 2026, 17:54 UTC

• Apache-2.0

The library provides a collection of WebGPU kernels for machine learning workloads. The collection contains 207 kernels and is licensed under Apache-2.0.

huggingface.cohuggingface-blog1 Sept 2026, 00:00 UTC

It provides a flashcard application for Android. It runs on Android devices.

github.comhn-frontpage1 Sept 2026, 10:11 UTC

• 770B params
• 1M context

The model produces text output from text input. The model features 770B total parameters, 49B active parameters, a 1M token context window, and a size of 1.56TB.

simonwillison.netsimonwillison29 Aug 2026, 23:53 UTC

The platform hosts open-weight AI models and benchmarks for developers. The source text contains no hardware or performance specifications.

techcrunch.comtechcrunch-ai28 Aug 2026, 18:19 UTC

• 125B params
• 262K context
• GGUF

The model produces reasoning outputs for a 262K context window. The model requires 102GB of total RAM and VRAM.

github.comunsloth-releases27 Aug 2026, 14:48 UTC

• Apache-2.0
• 20s clips
• INT8

Produces text-to-speech audio in English, Portuguese, French, and German. Uses INT8 quantization for 20-second clips under an Apache-2.0 license.

huggingface.cohf-models-new-tts29 Aug 2026, 22:14 UTC

• 0.6B params
• Apache-2.0
• 3s clips
• 10 languages

The model produces text-to-speech audio in ten languages. The model uses 0.6B parameters and FP4 weights.

huggingface.cohf-models-new-tts29 Aug 2026, 20:54 UTC

• CC-BY-NC

The model produces Japanese text-to-speech audio for the character Chi from the anime Chobits. It is a GPT-SoVITS v2Pro fine-tuned model trained on 483 audio samples.

huggingface.cohf-models-new-tts29 Aug 2026, 00:36 UTC

Produces Khakas language speech from text input without requiring manual stress marks. Utilizes the Silero v5-cis base engine for processing.

huggingface.cohf-models-new-tts28 Aug 2026, 11:01 UTC

There's a lot of capital pouring into the business of giving models away. No specs stated in the source — open the link before quoting numbers.

techcrunch.comtechcrunch-ai28 Aug 2026, 18:19 UTC

• 125B params
• 262K context
• GGUF

The model produces multimodal reasoning. It requires 102GB of total memory.

github.comunsloth-releases27 Aug 2026, 14:48 UTC

• 125B params
• GGUF

The model produces multimodal outputs using a Mixture of Experts architecture. The model features 125B parameters and is available in GGUF format.

simonwillison.netsimonwillison26 Aug 2026, 23:52 UTC

• 0.6B params
• Apache-2.0
• 3s clips
• 10 languages

It produces speech for 10 languages from text input. The model has 0.6B parameters and uses FP4 weights.

huggingface.cohf-models-new-tts29 Aug 2026, 20:54 UTC

Produces speech and audio from text. No hardware requirements are listed.

huggingface.cohf-models-new-tts29 Aug 2026, 13:17 UTC

• CC-BY-NC

Produces Japanese text-to-speech audio for the character Chi from the anime Chobits. This is a fine-tuned GPT-SoVITS v2Pro model based on 483 audio clips.

huggingface.cohf-models-new-tts29 Aug 2026, 00:36 UTC

Produces text-to-speech audio and voice. Requires GPT-SoVITS v2 and matching GPT and SoVITS weight files.

huggingface.cohf-models-new-tts27 Aug 2026, 15:48 UTC

• 198B params

The model produces text using gated residual, sparse attention, and per-layer embedding. It requires hardware to support 198 billion parameters.

github.comtransformers-releases26 Aug 2026, 15:03 UTC

• 27B params

The model produces text. The model has 27B parameters.

blog.eleuther.aieleutherai25 Aug 2026, 16:35 UTC

• 125B params
• 262K context
• GGUF

Produces multimodal reasoning outputs. 125B parameter model requiring 102GB total memory.

github.comunsloth-releases27 Aug 2026, 14:48 UTC

• 125B params
• GGUF

This model produces multimodal output. It features 125B parameters with 6B active and is available in GGUF format.

simonwillison.netsimonwillison26 Aug 2026, 23:52 UTC

Ox Alpha produces code, handles sustained agentic work, and manages production workloads with complex reasoning and visual context. The model is released with open weights.

techcrunch.comtechcrunch-ai26 Aug 2026, 14:19 UTC

• 3B params

Granite 4.2 is a decoder-only language model with a 128,000-token context window. The model is available in 3B, 8B, and 30B parameter variants.

arstechnica.comarstechnica-ai26 Aug 2026, 11:10 UTC

• 27B params

Provides bug fixes, auto compaction, LAN access, and improved streaming performance. The model has 27B parameters.

github.comunsloth-releases25 Aug 2026, 12:14 UTC

The model generates text-to-speech audio in English and Chinese. The model runs using the transformers library with PyTorch and CUDA support.

huggingface.cohf-models-trending25 Aug 2026, 16:39 UTC

• Apache-2.0

It provides retrainable robot behaviors such as walking, sitting, and grabbing trained in physics simulations for deployment on hardware. The project is licensed under Apache-2.0.

pollen-robotics.comhn-frontpage27 Aug 2026, 10:57 UTC

• 198B params

Produces text and multimodal content. Requires hardware capable of supporting 198 billion parameters.

github.comtransformers-releases26 Aug 2026, 15:03 UTC

• 27B params

Identifies whether an AI agent produces false information or conceals known facts. Runs on a 27B parameter model.

blog.eleuther.aieleutherai25 Aug 2026, 16:35 UTC

• 125B params
• GGUF

This multimodal MoE model serves as a preview of the architecture used in Qwen4. The model has 125B parameters in GGUF format.

simonwillison.netsimonwillison26 Aug 2026, 23:52 UTC

The model produces reasoning for coding, sustained agentic work, and workflows combining text with visual context. It runs on released open weights.

techcrunch.comtechcrunch-ai26 Aug 2026, 14:19 UTC

• 3B params

This decoder-only text model features a 128,000-token context window and supports tool use. The model is available in 3B, 8B, and 30B parameter variants.

arstechnica.comarstechnica-ai26 Aug 2026, 11:10 UTC

• 27B params

The update provides bug fixes, auto compaction for long conversations, LAN remote access, and Qwen3.8-27B GGUF models. The 27B parameter model runs on hardware supporting MLX, RDNA GPUs, and custom llama.cpp builds.

github.comunsloth-releases25 Aug 2026, 12:14 UTC

• 8B params
• Apache-2.0
• 4K

developerjeremylive published SenseNova-U1.5-8B-MoT-etheroi to the Hugging Face hub for any-to-any. Specs: 8B params · Apache-2.0 · 4K.

huggingface.cohf-models-new-multimodal25 Aug 2026, 06:22 UTC

• Apache-2.0
• 20s clips

Produces text-to-speech voice and audio. Runs on an RK3576 board via a Docker container.

huggingface.cohf-models-new-tts26 Aug 2026, 07:00 UTC

• 4B params
• non-commercial licence

The model produces text-to-speech audio and voice cloning. The model contains 4B parameters and utilizes the transformers library.

huggingface.cohf-models-new-tts26 Aug 2026, 05:43 UTC

It produces speech and audio from text with support for voice cloning and design in English and Chinese. The model runs using the transformers and pytorch libraries on CUDA hardware.

huggingface.cohf-models-trending25 Aug 2026, 16:39 UTC

The system provides a virtual executive team for strategy, finance, HR, legal, operations, marketing, and product roadmap. No hardware or software requirements are listed in the source.

github.comhn-frontpage27 Aug 2026, 01:46 UTC

It provides point-to-point WireGuard-encrypted tunnels between machines using a token-based exchange. It runs as a command-line tool or as an importable Go library.

github.comhn-frontpage26 Aug 2026, 17:42 UTC

• non-commercial licence

Provides offline map access for navigation in areas without cellular signal. Licensed for non-commercial use.

hotosm.orghn-frontpage26 Aug 2026, 17:20 UTC

• 27B params

The update adds bug fixes for MLX and AMD hardware, LAN access, auto-compaction, and custom llama.cpp build support. The model utilizes 27B parameters.

github.comunsloth-releases25 Aug 2026, 12:14 UTC

• 8B params
• Apache-2.0
• 4K

The model provides any-to-any multimodal capabilities including text, image generation, and image editing. The model has 8B parameters and uses the transformers library.

huggingface.cohf-models-new-multimodal25 Aug 2026, 06:22 UTC

• 2B params
• Apache-2.0
• 48 kHz
• 30 languages

The model produces 48kHz audio in 30 languages using a tokenizer-free, diffusion autoregressive architecture. The model has 2B parameters and is released under an Apache-2.0 license.

huggingface.cohf-models-new-tts25 Aug 2026, 13:47 UTC

• 0.1B params
• 44.1 kHz

The model produces text-to-speech audio and zero-shot voice cloning in multiple languages. The model uses 0.1B parameters and a 44.1 kHz sampling rate.

huggingface.cohf-models-new-tts25 Aug 2026, 10:57 UTC

• 0.1B params
• 44.1 kHz

The model produces zero-shot text-to-speech audio in multiple languages. The model has 0.1 billion parameters and a 44.1 kHz sample rate.

huggingface.cohf-models-new-tts25 Aug 2026, 10:00 UTC

Produces text-to-speech audio for the Yoruba language. Requires gated access to download VITS weights in safetensors format.

huggingface.cohf-models-new-tts24 Aug 2026, 15:15 UTC

• 32 kHz

Converts text into voice outputs for Chinese, Japanese, and English. Operates on the GPT-SoVITS library with a 32 kHz sample rate.

huggingface.cohf-models-new-tts24 Aug 2026, 14:07 UTC

• 8B params
• Apache-2.0

The model produces speech and audio from text input. The model has 8B parameters.

huggingface.cohf-models-new-tts24 Aug 2026, 12:19 UTC

• 26B total / 4B active
• 26B params
• 128K context
• Apache-2.0
• 140 languages

The model performs any-to-any tasks across 140 languages with a 128K context window. The model has 26 billion total parameters and 4 billion active parameters.

huggingface.cohf-models-new-multimodal22 Aug 2026, 02:20 UTC

• Apache-2.0

Produces Amharic text using four 50M-parameter model architectures. Requires the transformers library and follows an Apache-2.0 license.

huggingface.cohf-models-amharic21 Aug 2026, 15:48 UTC

• Apache-2.0

Produces Japanese text-to-speech audio using the GPT-SoVITS V2 model. Requires the GPT-SoVITS repository and weight files of 155 MB and 75 MB.

huggingface.cohf-models-new-tts23 Aug 2026, 17:31 UTC

The model produces speech from text and handles image-to-text and text-to-image tasks. The model requires the pytorch library.

huggingface.cohf-models-new-tts21 Aug 2026, 05:17 UTC

• Apache-2.0

The model performs any-to-any tasks. The model runs using the diffusers library and safetensors weights.

huggingface.cohf-models-new-multimodal21 Aug 2026, 15:10 UTC

• 8B total / 1B active
• 2.6B params
• GGUF

- Faster inference: up to 3.18 throughput improvement on a GPU and up to 2.87x on-device. Specs: 8B total / 1B active · 2.6B params · GGUF.

huggingface.cohuggingface-blog20 Aug 2026, 16:52 UTC

• 27B params
• GGUF

The update provides auto-compaction for conversations exceeding context limits, LAN access, and improved streaming performance. It runs on 27B parameter models in GGUF format.

github.comunsloth-releases20 Aug 2026, 18:05 UTC

• 27B params
• Apache-2.0

Mojo🔥 is now open source The Mojo programming language has been promising an open source release since May 2023 . Specs: 27B params · Apache-2.0.

simonwillison.netsimonwillison18 Aug 2026, 21:39 UTC

• 7B params
• 6s clips
• 24 kHz
• 17 languages

The model produces speech in 17 languages and clones voices from 6-second audio clips. The model features 7B parameters and produces 24 kHz audio.

huggingface.cohf-models-new-tts21 Aug 2026, 16:04 UTC

Produces speech from text and performs image-to-text and text-to-image tasks. Requires PyTorch and external source code from a separate GitHub repository for inference.

huggingface.cohf-models-new-tts21 Aug 2026, 05:17 UTC

The model produces text-to-speech audio and instant voice cloning. The system requires OpenVoice V1 and V2 base speaker and tone-color-converter checkpoints.

huggingface.cohf-models-new-tts20 Aug 2026, 22:39 UTC

• Apache-2.0
• 16 kHz

Produces Chinese text-to-speech with independent speaker and accent references. Uses a VITS-style acoustic model and HiFiGAN vocoder at 16 kHz.

huggingface.cohf-models-new-tts20 Aug 2026, 17:50 UTC

• 5B params
• Apache-2.0

It produces text-to-speech audio with controllable frame rates. It has 5B parameters and runs with the transformers library.

huggingface.cohf-models-new-tts20 Aug 2026, 04:15 UTC

• 7B params
• Apache-2.0

The model produces speech audio from text input. The model has 7B parameters.

huggingface.cohf-models-new-tts20 Aug 2026, 03:34 UTC

• 0.1B params
• 44.1 kHz

The model produces text-to-speech audio for Chinese, English, German, Spanish, French, Italian, Japanese, and Korean. The model has 0.1B parameters and uses the transformers library.

huggingface.cohf-models-new-tts19 Aug 2026, 04:49 UTC

• Apache-2.0
• 4 kHz

Produces text-to-speech voice and audio. Operates at 4 kHz with a single-step time of 181 ms and a real-time factor of 0.3–1.1.

huggingface.cohf-models-new-tts19 Aug 2026, 04:20 UTC

• 90s clips

The model produces video with integrated stereo audio, dialogue, and foley in one pass. The model runs on local hardware or Comfy Cloud.

blog.comfy.orgcomfyui-blog19 Aug 2026, 20:39 UTC

The tool measures average and extreme fidelity loss for vendor-hosted LLM APIs using repeated-request and long-sequence analyses. The audit functions as a black-box system requiring no probability information from the target API.

huggingface.cohf-daily-papers16 Aug 2026, 20:00 UTC

• 27B params
• Apache-2.0

It is a programming language optimized for GPU programming. The compiler and toolchain are released under an Apache-2.0 license.

simonwillison.netsimonwillison18 Aug 2026, 21:39 UTC

• 27B params
• Apache-2.0
• GGUF

The model functions as a vision-capable large language model. The model has 27B parameters and is available as a 17GB Q4_K_M quantized GGUF build.

simonwillison.netsimonwillison16 Aug 2026, 22:00 UTC

The Qwen 3.8 2.4T model produces a Call of Duty clone from a single prompt. The model requires a B200 cluster to run.

reddit.comreddit-localllama18 Aug 2026, 17:53 UTC

• FP8

The model produces text-to-speech for Amharic and Geez. The model supports FP8.

huggingface.cohf-models-amharic18 Aug 2026, 15:31 UTC

Produces Mandarin text-to-speech audio using a diffusion-based acoustic model. Requires the PyTorch library for execution.

huggingface.cohf-models-new-tts16 Aug 2026, 09:12 UTC

The system allows users to build, edit, and run ComfyUI workflows via plain language and connects local and cloud environments to MCP clients. The system requires a local ComfyUI installation, a Comfy account, and an MCP-compatible client such as Claude, Cursor, or Codex.

blog.comfy.orgcomfyui-blog18 Aug 2026, 16:46 UTC

The model produces coordinate-aware token-space representations to reconstruct spatio-temporal queries from video data. It requires one feed-forward pass and a shared decoder.

arxiv.orgarxiv-cv17 Aug 2026, 04:00 UTC

• 70B params

It produces Arabic-centric text and performs on culturally grounded benchmarks. The model comes in 70B and 8B parameter versions.

arxiv.orgarxiv-cl17 Aug 2026, 04:00 UTC

• 27B params
• Apache-2.0
• GGUF

Produces text and vision-capable outputs. Runs on a 128GB M5 Max MacBook Pro or NVIDIA DGX Spark using a 17GB Q4_K_M quantized build.

simonwillison.netsimonwillison16 Aug 2026, 22:00 UTC

• Apache-2.0

The model performs named entity recognition in Amharic. It requires the transformers library and Hugging Face model weights.

huggingface.cohf-models-amharic17 Aug 2026, 15:31 UTC

Produces text-to-speech audio in English and several Indian languages. No hardware requirements are provided.

huggingface.cohf-models-new-tts15 Aug 2026, 09:38 UTC

• 120B params

The two leading labs are already feeling extremely threatened by Qwen & friends, however I think there are tons of enterprises and organizations in the West that don't feel comfortable using Chinese models. Specs: 120B params.

reddit.comreddit-localllama15 Aug 2026, 21:53 UTC

• 754B params

Models and datasets on HF hub are growing on a daily basis. Specs: 754B params.

huggingface.cohuggingface-blog14 Aug 2026, 00:00 UTC

• Apache-2.0

The system records robot demonstrations, trains policies on the data from Hugging Face Hub, and deploys them to hardware. Licensed under Apache-2.0.

huggingface.cohuggingface-blog13 Aug 2026, 17:16 UTC

The model produces text responses through an LLM interface. The weights are available for 1.7 trillion parameters and require 893 GB of storage.

simonwillison.netsimonwillison12 Aug 2026, 23:59 UTC

• 30B params

Produces text output via a mixture-of-experts architecture for agentic tasks. Requires 30B total parameters and 3B active parameters.

github.comollama-releases11 Aug 2026, 13:24 UTC

• 280B params

Produces text outputs based on text, image, video, and audio inputs. Mixture-of-Experts model with 280B total parameters, 16B activated parameters, and a 512K token context length.

reddit.comreddit-localllama13 Aug 2026, 21:46 UTC

• MIT

DeepSeek Harness is an agent framework where components like models, tools, and sandboxes are implemented as plugins. The code is released under the MIT license.

reddit.comreddit-localllama13 Aug 2026, 13:32 UTC

• 8B params
• 32 kHz

Generates full songs up to five minutes in length based on lyrics and a description of the sound. The model contains 8B parameters and produces 32 kHz stereo audio.

blog.comfy.orgcomfyui-blog13 Aug 2026, 20:41 UTC

• 1M context
• 8K
• FP8

Highlights 582 PRs from 194 contributors. Specs: 1M context · 8K · FP8.

github.comsglang-releases11 Aug 2026, 03:17 UTC

• Apache-2.0

The system manages recording robot demonstrations, training policies on them using LeRobot data formats, and deploying the models to hardware. This software is licensed under Apache-2.0.

huggingface.cohuggingface-blog13 Aug 2026, 17:16 UTC

• 8-BIT

Produces embedding vectors for earth observation data analysis. Available in 8-bit format.

huggingface.cohuggingface-blog12 Aug 2026, 16:14 UTC

The model powers coding agent applications and long-running personal assistants. It runs on all platforms including Apple Silicon, NVIDIA, and AMD hardware.

github.comollama-releases11 Aug 2026, 06:39 UTC

• 30B params
• Apache-2.0

Produces text for agentic use cases, coding, document analysis, and personal assistants. 30B parameter dense model with a 2B ViT-style vision encoder and 28B text decoder under Apache-2.0 license.

github.comtransformers-releases10 Aug 2026, 10:28 UTC

The system identifies differences between benchmark scores and deployment performance by masking primary candidate tokens at word boundaries during runtime. No specific technical specifications are provided in the source text.

huggingface.cohf-daily-papers9 Aug 2026, 20:00 UTC

• 30B params

Produces text outputs from a mixture-of-experts model for agent workflows. Requires 30B total parameters with 3B active parameters.

github.comollama-releases11 Aug 2026, 13:24 UTC

• 104B params
• 1M context
• GGUF

Produces text and long context outputs. Requires 104B parameters, up to 162TB for lossless inference, and compatible with GGUF format.

github.comunsloth-releases10 Aug 2026, 04:02 UTC

• 12 languages

The model generates speech for multiple languages. It supports 12 languages and requires open weights deployment.

huggingface.cohuggingface-blog10 Aug 2026, 16:25 UTC

• 12B params
• 720p
• 4K
• 20s clips
• 50 fps

The model generates multi-shot videos with synchronized audio and frames at a rate of 50 fps. It uses 12B parameters and runs on local GPUs to output video in 720p or 4K resolutions.

blog.comfy.orgcomfyui-blog12 Aug 2026, 04:42 UTC

Generates video from images and text. Requires access to the Hugging Face repository via a gated license.

huggingface.cohf-models-lightricks10 Aug 2026, 11:03 UTC

• 1M context
• 8K
• FP8

Kimi K3 produces multimodal content with a 1M-token context and MoonViT3d vision tower. The model uses 2.8T parameters, FP8 quantization, and supports SGLang serving with DCP and speculative decoding.

github.comsglang-releases11 Aug 2026, 03:17 UTC

Provides a graphical interface for stable diffusion and other image generation models. Requires PyTorch 2.7 or higher to run.

github.comcomfyui-releases11 Aug 2026, 21:56 UTC

• 30B params
• GGUF

Unsloth Desktop is here! Specs: 30B params · GGUF.

github.comunsloth-releases11 Aug 2026, 15:57 UTC

• 30B params

The model generates text using a mixture-of-experts architecture with 3B active parameters. It requires an infrastructure supporting the NVIDIA NemoClaw stack and is available via Ollama.

github.comollama-releases11 Aug 2026, 13:24 UTC

• 30B params
• Apache-2.0

The model produces end-to-end agentic task completion, tool use, and multi-step reasoning. The model has 30 billion parameters and is under an Apache-2.0 license.

simonwillison.netsimonwillison10 Aug 2026, 23:56 UTC

• Apache-2.0

Muse Glimmer produces text outputs for large language model tasks. The model has 30 billion parameters and a 128,000-token context window.

arstechnica.comarstechnica-ai10 Aug 2026, 22:13 UTC

• 1.1B params
• GGUF

This system enables llama.cpp to utilize advanced Metal kernels for inference on macOS virtual machines. Requires a host with Apple Silicon and the Lume virtualization stack.

github.comhn-frontpage11 Aug 2026, 14:50 UTC

• 1M context
• 8K
• FP8

SGLang supports Kimi K3 and MiniMax-H3 models for multimodal, video, and audio generation. The system supports 1M context, 8K resolution, and FP8 precision.

github.comsglang-releases11 Aug 2026, 03:17 UTC

Muse Glimmer produces coding agent outputs and responses for personal assistant frameworks. Requires Ollama software and runs on NVIDIA, AMD, and Apple Silicon hardware.

github.comollama-releases11 Aug 2026, 06:39 UTC

• 30B params

To celebrate, we are shipping with Meta day-0 support in transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries. Specs: 30B params.

huggingface.cohuggingface-blog10 Aug 2026, 00:00 UTC

• 30B params
• Apache-2.0

The model produces text and processes vision input for coding, document analysis, and personal assistant tasks. It is a 30B parameter dense model with a 2B ViT-style encoder and a 28B parameter text decoder.

github.comtransformers-releases10 Aug 2026, 10:28 UTC

• 30B params

Muse Glimmer is a multimodal model for agent workloads and coding applications. The model has 30B parameters.

github.comollama-releases10 Aug 2026, 11:13 UTC

WorldClaw generates 3D content. No specific hardware or software requirements are listed.

reddit.comreddit-localllama9 Aug 2026, 11:42 UTC

• 0.8B params

It identifies steerable semantic concept directions to control behavior in multiple independently trained large language models. The study tests five open-weight models with parameter counts from 0.8B to 8B.

arxiv.orgarxiv-cl7 Aug 2026, 04:00 UTC

• 750B params
• Apache-2.0

Produces multilingual text generation across ten languages with a 256K token context length. Operates as a Mixture-of-Experts model with 750B total parameters and 37B activated per token.

arxiv.orgarxiv-cl6 Aug 2026, 04:00 UTC

Ollama provides a local runner for large language models with updated TUI features and support for the Qwen3.5 model. The software runs on Ollama v0.32.6.

github.comollama-releases5 Aug 2026, 19:46 UTC

• 20s clips

Produces video clips of 20 seconds. Requires ComfyUI and Partner Nodes.

blog.comfy.orgcomfyui-blog5 Aug 2026, 19:31 UTC