LangChain just released OpenWiki, an open-source CLI agent that automatically generates and maintains documentation for codebases. Instead of cramming an entire repo into CLAUDE.md, it builds a structured wiki and syncs changes via GitHub Actions, giving coding agents durable context without the bloat.
Microsoft just open-sourced SkillOpt, a framework that optimizes agent skill files like deep-learning parameters without altering underlying model weights. By using text-space forward passes, reflective edits, and strict validation gating, it boosted GPT-5.5 benchmark scores by 23.5 points. Why it matters: It replaces messy manual prompt hacking with a reproducible, controlled training loop.
Researchers just open-sourced OVOW, a training-free system that turns a single video into simulation-ready 4D mesh scenes with separated objects. By chaining models like Qwen3-VL and SAM 3, it reconstructs motions and deformations directly for physics simulators. It is available for non-commercial research and recommends at least 40GB of VRAM.
Meituan just open-sourced LongCat-2.0, a massive 1.6-trillion-parameter MoE model that ditches Nvidia hardware entirely. The pretraining run spanned over 35 trillion tokens on dedicated AI ASIC superclusters without irrecoverable loss spikes, matching frontier models across coding and agent tasks. Why it matters: It proves cutting-edge AI can be trained at scale on alternative chips.
Microsoft just open-sourced Resource2Skill, a framework that distills human tutorial videos, articles, and code into executable agent skills. Instead of manual prompt writing, it builds a multimodal skill library hooked up to MCP servers, enabling agents to generate real Web pages, PowerPoint decks, and Blender scenes.
Einsia AI's Navers Lab just open-sourced BrowserBC, a tool that acts like an AI-native macro recorder. Instead of brittle pixel coordinates, it records your browser actions and distills them into plain-text skill cards that any web agent can read, understand, and execute autonomously.
Ornith just introduced Ornith-1.0, an open-source model family designed specifically for agentic coding across sizes from 9B Dense to 397B MoE. The key breakthrough is self-scaffolding: the model learns to author its own task workflows rather than relying on human-engineered harnesses, matching top frontier models on SWE-bench Verified. Weights and edge-ready 21GB GGUF variants are live on Hugging Face.
Meta just open-sourced Astryx, the internal design system that has powered over 13,000 of its apps over the past eight years. Built on React 19, it ships 150+ accessible components, ready-made themes, and a CLI designed so human engineers and AI coding agents can build from the exact same conventions.
Nous Research just added a /learn command to Hermes Agent, letting it distill any directory, documentation URL, or recent workflow directly into a reusable skill file. It skips heavy background engines entirely, using the agent's own file and browsing tools so it works seamlessly across local and remote setups.
Krea just released open weights for two new image models: Krea 2 Raw and Krea 2 Turbo. Raw is an undistilled, mid-training checkpoint built specifically for developers to fine-tune, while Turbo is a distilled, fast-inference version offering broad aesthetic variety. Why it matters: developers can now build and customize Krea's generation tech on their own hardware.
HKUST just open-sourced Domain Shuttle, a subject-driven video generation tool that preserves reference character consistency even across mixed artistic styles. It lets creators combine live-action, 2D anime, and 3D characters in a single scene while transforming styles using text prompts. The code and 35GB model based on Wan 2.2 are available on GitHub, requiring a high-end GPU.
InternScience just open-sourced Agents-A1, a 35B mixture-of-experts model built specifically for multi-step agentic workflows and tool use. Despite its compact footprint, it rivals frontier giants like GPT-5.5 on science and agent benchmarks, with quantized versions small enough to run locally on consumer hardware.
Liquid AI just rolled out two 350M-parameter retrieval models, LFM2.5-ColBERT and LFM2.5-Embedding, bringing fast multilingual search across 11 languages to edge devices and CPUs. The twin models tackle both single-vector speed and token-level accuracy, outperforming rivals on NanoBEIR and MKQA benchmarks. Why it matters: developers can now run high-accuracy cross-lingual search locally via GGUF without blowing up compute budgets.
Ai2 just open-sourced MolmoMotion, a multimodal model that predicts how 3D points on objects will move based on video frames and text prompts. Built on Molmo 2, it boosts simulated robotic pick-and-place success from 56% to 76.3% and steers controllable video generation. Ai2 also released an open dataset of 1.16M video trajectories and a new benchmark.
PermaVid just open-sourced a video editing framework that solves one of generative video's biggest headaches: consistency. By splitting context into two separate memory banks for RGB appearance and 3D depth geometry, it keeps scene edits stable across time without distorting structure. Why it matters: It stops AI from forgetting edits, making iterative long-form video editing actually viable.
Moonshot AI just open-sourced Kimi K2.7 Code, a massive 1-trillion parameter MoE model with 32B active parameters. It slashes reasoning token usage by roughly 30% while closing the gap on top proprietary frontier models like GPT-5.5 and Opus 4.8 in long-horizon programming tasks. Full 600GB weights are now available on Hugging Face.
Zyphra just released ZONOS2, an open-source TTS model with 8B total and 900M active parameters powered by MoE. Trained on over 6M hours of multilingual speech, it delivers low-latency 44.1 kHz voice cloning under an Apache 2.0 license, with ready-to-run local inference and a free web playground.
Google just open-sourced DiffusionGemma, an experimental 26B MoE model that ditches token-by-token generation for text diffusion, stamping out 256 tokens in parallel. It delivers up to 4x faster GPU inference, clocking over 1,000 tokens per second on an H100 and fitting into 18GB VRAM when quantized. Why it matters: It drastically cuts latency for local inline editing and code infilling, though Google notes raw output quality still trails standard Gemma 4.
