New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
NVIDIA just released Nemotron 3 Embed, a trio of open-weight models aimed at production RAG and agentic workflows. The 8B flagship took first place on the RTEB multilingual benchmark, while a 1B variant gets native NVFP4 acceleration for Blackwell chips. Beyond 32k context and code retrieval, the key hook is efficiency: cleaner context retrieval directly cuts down multi-turn agent reasoning costs.
Bolt just open-sourced Bolt Slides, a presentation builder that turns static decks into living web applications. Instead of exporting dead PDFs, users can embed live data, 3D models, and clickable prototypes directly into slide links. It is available to all Bolt.new users and can also be run locally with agents like Cursor or Claude Code.
xAI just open-sourced Grok Build, releasing the harness behind its coding agent and terminal UI on GitHub. The release details how it handles context assembly and tool dispatch, while offering plug-in support for MCP servers and subagents. Best of all, developers can compile it themselves and run it fully local-first with their own inference.
OpenMOSS just open-sourced MOSS-VL-Realtime, an 11B parameter vision-language model tailored for live streaming video. Instead of waiting for an entire clip to load before answering, it processes incoming frames on the fly, allows users to interrupt with questions at any timestamp, and stays silent when visual evidence is insufficient.
PrismML just released Bonsai 27B, an open-source multimodal model based on Qwen3.6 that fits onto consumer phones and laptops. Its extreme 1-bit binary variant shrinks the memory footprint to just 3.9GB—roughly 14x smaller than standard FP16—while hitting 11 tokens per second on an iPhone 17 Pro. Both the 1-bit and 1.58-bit ternary models are freely downloadable under the Apache 2.0 license.
Alibaba open-sourced Wan-Animate-2, an end-to-end character animation model that animates still photos using driving videos. It replicates body movements, delicate finger gestures, and facial expressions—even on non-human characters and multi-person scenes—while supporting camera viewpoint controls.
Developer Syed Subhan just open-sourced Amicro, a collection of hardware-accelerated React micro-interactions and transitions powered by Motion and Tailwind CSS. It works directly with the shadcn CLI to drop TSX components straight into your codebase with zero runtime overhead.
Open-source matting model Lucida just launched to fix where typical removers fail, cleanly isolating semi-transparent glass, camouflage, drop shadows, and fine hair. At under 900MB on an MIT license, it runs locally on consumer GPUs and beats commercial baselines across several edge cases.
Nvidia just updated its open-source PID decoder to version 1.5, bringing sharper detail and cleaner color fidelity without grid artifacts. It turns 512×512 latents straight into 2K or 4K renders on consumer GPUs in mere seconds. Why it matters: it bypasses slow, two-step upscaling workflows and works out of the box with models like FLUX and Qwen-Image.
Mirelo and Kyutai just open-sourced MuScriptor, a new audio-to-MIDI model that handles full, complex music mixes. Instead of tripping over overlapping instruments like older single-track tools, it transcribes entire songs into separate MIDI tracks for vocals, drums, bass, and keys all at once. It's basically a dream shortcut for producers and learners looking to rip clean tracks straight into their DAW.
Ant Group's Robbyant just open-sourced LingBot-World 2.0, an interactive world model that streams real-time 720p video at 60 fps. It pairs a causal pretraining approach with dual pilot-and-director agents to handle character actions and dynamic scene building over extended interactions. Weights for 1.3B and 14B models are up on Hugging Face.
Qualcomm researchers just open-sourced MobileWan, compressing Alibaba's 5B-parameter video model to run entirely on commercial smartphones. By pruning attention heads and slashing diffusion to three steps, it generates a 5-second 480p clip in about 20 seconds without touching the cloud.
Researchers just released MIRA, an interactive multiplayer world model that simulates Rocket League without any physics or rendering engine. Running on a single B200 GPU, the 5B diffusion model generates consistent 20 fps views responding to real-time inputs from four players, and the code and dataset are fully open-sourced.
Tencent just open-sourced Hy3, a 295B-parameter Mixture-of-Experts model that activates 21B parameters per token. Tuned for agentic workflows and coding, it boasts a 256K context window and cuts hallucination rates down to 5.4% in internal evaluations. It offers developers a cost-effective, Apache 2.0-licensed alternative to massive proprietary frontier models.
Developer Steven Chong just open-sourced pxpipe, a local proxy that squeezes Claude Code context and system prompts into dense PNG images. Because Anthropic prices images by pixel dimensions rather than characters, users can cut token costs by up to 70%, though vision decoding adds slight latency and string loss.
Researchers just released Mr. Flow, a training-free framework that accelerates text-to-image models by 9x to 21x. It drafts quickly at low resolution, upscales with subtle noise, and refines high-frequency details in the final step, slashing wait times to around one second without hurting quality.
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.
