Kyutai Ships OVIE-512: One Image, One New View, 41.6 Milliseconds
A single-frame novel-view-synthesis model, MIT-licensed, with no diffusion steps and no ComfyUI node. Fast, small, and not a video generator.
Category
AI News
XiaomiMiMo Ships a 309B MoE Model Targeted at One Specific Agentic Bug
MiMo-V2.6-Flash-MOPD is the MOPD-trained successor to the -RL checkpoint. Its stated purpose is not a benchmark bump — it is fixing tool-call repetition in agentic loops.
Ming-UniVision-16B-A3B: one model for understanding, generating, and editing images
inclusionAI's unified multimodal model drops discrete quantization and modality-specific heads, but the two-turn training ceiling and custom inference path mean small studios are watching, not shipping.
TencentARC Ships WorldCrafter-Base: An Image-to-Video Model That Needs a Sibling to Run
A two-checkpoint video generation model with a 3D-aware memory architecture, no stated licence, no ComfyUI node, and no VRAM floor. The lab put it on the hub; the rest is on you.
Xiaomi's MiMo-V2.6-Flash-RL: 309B Parameters, 15B Active, MIT Licensed
A four-modality, 1M-context agent model that costs like a 15B model per token — and the first time a hardware company has put that combination under MIT.
Shanghai AI Lab Ships Atria Dawn Preview: A 744B Agentic MoE Under an MIT Licence
Three quantized variants, a 256K context window, and an MIT licence that makes the orchestration story cleaner than the hardware story.
IBM's 385M-parameter time-series model beats the bigger ones
Granite Time Series PatchTST-FM-r2 tops the permissively-licensed zero-shot tier on GIFT-Eval. The conformer swap and the 99-quantile head matter more than the leaderboard position.
NeoMME matches a retriever 14× its size at 260M parameters
H Company pretrained a multimodal encoder from scratch instead of repurposing a generative VLM, and released the weights under Apache-2.0.
Microsoft ships a streaming ASR model that already knows who is talking
VibeVoice-ASR-Streaming-7B folds speaker attribution into the transcription pass, drops the diarisation step, and is MIT-licensed. No benchmarks yet.
IFM Ships Six Open LLMs and the Entire Lab Notebook Behind Them
K2 Horizon covers 0.9B to 375B under Apache-2.0, but the real release is the training data, intermediate checkpoints, and agentic post-training recipes published alongside.
Tencent Ships a 770B MoE Text Model Under Apache-2.0 — and the Benchmarks Are Its Own
Hy4-preview lands on Hugging Face with 49B activated parameters, a 1M-token window, and a licence that lets you ship. The benchmark data, however, is internal, and the memory math is the number that should matter more.
Gemini Omni 1.1 Flash: Scene Extension, Keyframe Control, and a 360p Draft Tier
Google DeepMind shipped a generative video API update that jumps the scene-continuity context window tenfold, adds a one-third-cost draft tier, and introduces first-and-last-frame interpolation — all cloud-only, all on the Gemini API.
ComfyUI Open-Sources Its MCP Server for Local Use
The MCP server that once only talked to a cloud instance now runs beside your local ComfyUI, reads your GPU, and can reach into Blender and DaVinci on the same file system. One agent session covers both paths.
AWS Strands Robots SDK Bundles LeRobot Streaming Into Composable Agent Tools
How a new AWS robotics SDK handles cloud storage, frame decoding, and local inference for continuous training loops.
LTX-2.5 Gets Day-0 ComfyUI Support With Dynamic Compute And Multi-Shot Sequencing
Open weights ship alongside a required ComfyUI version bump. Studio implications for local batching and fine-tuning.
Google AMIE (Video) Decouples Medical AI Consultations via Async Multi-Agent Routing
The update achieves expert-level diagnostic performance against physicians by splitting perception and reasoning into parallel workers, validating a latency-preserving architecture for local AV stacks.
ComfyUI v0.32.0 Forces PyTorch Upgrade and Ships Official Partner Nodes
The latest runtime release drops legacy torch support, closes H3 memory leaks, and routes Qwen-Image and Grok directly through core.
Muse Glimmer Targets Consumer VRAM With a 30B Multimodal Stack
Meta Superintelligence Labs ships a dense VLM with hybrid attention, DFlash drafting, and day-zero runtime support. Offline video ingestion is feasible on single GPUs, but integration gaps remain.
Selective Inpainting with Tailored Noise Fixes Diffusion Augmentation for Cross-Domain Few-Shot Detection
A new framework isolates visual and semantic domain gaps, enabling diffusion models to generate viable training frames without corrupting foreground structures.
MiniMax H3 drops open weights with native ComfyUI support
A local-first omni-modal video model targets the RTX 3060 and collapses five production steps into a single pass.