Addis PulseStudio

📡 Daily radar

Open-source model drops

What came out that you can actually download — grouped by what it makes, with the specs that decide whether it runs on your box. Every link goes to the weights, the repository or the paper, never to coverage of them. Pick a date to read an earlier day, or a modality to see only what it shipped.

9 open-weight releases in the last 72 hours, 432 across 56 archived days.more than a day old


Archive

Modality

9 releases on Tuesday, 27 October 2026

Tuesday, 27 October 2026

9 releases168 clusters from 178 items

Text / LLM6

MiniCPM-OCR

Text / LLM
  • 24 kHz

Extracts text from images such as scanned pages, screenshots, receipts, and tables. Uses a SigLIP vision encoder and a Qwen3.5 language model.

github.comtransformers-releases26 Oct 2026, 18:21 UTC

LBS-27B

Text / LLM
  • 27B params
  • Apache-2.0
  • 4K

The model produces text from image inputs for visual question answering and GUI grounding. The model has 27B parameters.

huggingface.cohf-models-inclusionai26 Oct 2026, 05:36 UTC

Ling-3.1-flash-base

Text / LLM
  • 560B params
  • FP8

Produces text for conversational interactions. Requires 560B parameters in FP8 precision.

huggingface.cohf-models-inclusionai25 Oct 2026, 02:15 UTC

Opera

Text / LLM

The feature allows the local execution of open-weight models and Apple's on-device model within the browser. It requires the Opera One, Opera GX, or Opera Neon browser software.

techcrunch.comtechcrunch-ai27 Oct 2026, 13:00 UTC

Alibaba

Text / LLM

The Qwen3.8-Flash-Next model generates text outputs for various instructions. The model is an open-weight architecture.

techcrunch.comtechcrunch-ai27 Oct 2026, 01:12 UTC

Qwen2.5-Omni-7B

Text / LLM
  • 7B params

The model performs any-to-any multimodal tasks. The model has 7B parameters and works with the transformers library.

huggingface.cohf-models-new-multimodal27 Oct 2026, 01:01 UTC

TTS / Voice1

f5-tts-russian-emotional

TTS / Voice
  • 10s clips
  • 24 kHz
  • 8GB VRAM
  • 8-BIT

Produces Russian text-to-speech and voice cloning. Requires 8GB VRAM in 8-bit to produce 10s clips at 24 kHz.

huggingface.cohf-models-new-tts27 Oct 2026, 08:12 UTC

Image / Other2

Energy-Steered

Image / Other

The model produces text tokens by refining latent states through energy-guided search. The source text does not provide specific hardware or computational requirements.

arxiv.orgarxiv-cl27 Oct 2026, 04:00 UTC

Unlimited-OCR-2

Image / Other

The model performs multi-language text extraction from images. It utilizes the transformers library and supports safetensors.

huggingface.cohf-models-baidu26 Oct 2026, 03:57 UTC

How this page is made

A scheduled run reads 178 items from the studio’s allowlisted feeds, groups them into 168 stories, and keeps the ones that read as an open-weight release: something announced, something downloadable, and something with a name. Each entry is tagged by vocabulary and summarised by a small open-weight model running on the studio’s own hardware.

Which means this page is generated, and no editor has read every line of it — it ships as part of a pull request a human merges, and the specs come from the source’s own words. Treat it as a tip-off, not as a fact-checked article: open the link before you quote a number or trust a licence. The writing is where a story gets read, researched and argued about by a person.

Nothing is overwritten. Each run leaves a dated copy of its report in the repository, which is what the archive rail at the top of the page walks — 56 days of drops, filterable by modality on any of them.

Run editorial-20261027T160538Z