New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Alibaba just open-sourced Qwen3.6-35B-A3B under Apache 2.0, a sparse MoE model packing 35B total parameters with only 3B active. Despite its tiny active footprint, it claims agentic coding performance on par with models 10x its active size, alongside multimodal reasoning and thinking modes.
Mozilla subsidiary MZLA just launched Thunderbolt, an open-source, self-hosted AI client designed to prevent vendor lock-in. Powered by deepset's Haystack RAG framework, it runs across desktop and mobile to handle search, chat, and multi-device automation. It is basically pitching data sovereignty for enterprise teams wary of cloud AI platforms.
Tencent's Hunyuan team just open-sourced HY-World 2.0, moving beyond video-only world models to produce actual, editable 3D assets. It transforms text, images, or video clips into meshes and 3D Gaussian Splats ready for Unity, Unreal Engine, and Isaac Sim with physics-aware navigation.
Nvidia just open-sourced Lyra 2.0, a lightweight 131MB framework that converts video clips into persistent, interactive 3D Gaussian splat environments. It eliminates temporal drift and spatial forgetting over long trajectories, allowing exported scenes to directly feed into Isaac Sim for training autonomous robots.
Spark just launched version 2.0, an open-source update built on Three.js that introduces a streamable Level of Detail (LoD) system for 3D Gaussian Splatting. Powered by WebGL2, it dynamically streams massive 100M+ splat worlds directly to browsers on everyday mobile phones and VR headsets, eliminating the need for beefy local hardware.
NVIDIA just launched Ising, an open-source family of quantum AI models designed for processor calibration and real-time error correction. Its calibration model uses vision language processing to cut setup time from days to hours, while the decoding models run up to 2.5x faster than pyMatching.
A new open-source algorithm called RotorQuant just outperformed Google's TurboQuant across every metric. By swapping heavy global transforms for compact Clifford rotators, it hits 10x KV cache compression while boosting prefill speed by 5.3x and decoding by 28%. Why it matters: massive 128K context windows that previously required multiple enterprise GPUs can now fit onto a single 24GB card.
MiniMax just open-sourced its new M2.7 coding model, clinching top scores on SWE-Pro (56.22%) and Terminal Bench 2 (57.0%). The weights are already live on Hugging Face alongside API access via the MiniMax platform, stepping up competition in open-source software engineering AI.
EditCrafter just dropped a tuning-free pipeline that lets pretrained diffusion models edit images up to 4K resolution. It preserves original structures and details without retraining, though it tends to oversaturate colors and demands 24GB of VRAM for full 4K processing.
AniGen just released an open-source framework that builds animatable 3D models with functional skeletons from a single photo. Instead of dealing with fragile post-processing, users get rigs naturally aligned to the mesh that are ready for animation software. Running it locally requires an NVIDIA GPU with at least 18GB of VRAM.
Researchers just released Numina, a training-free framework that forces text-to-video models to generate the exact number of objects requested in a prompt. By catching inconsistencies in attention heads and guiding regeneration, it improved counting accuracy on Wan2.1 by up to 7.4% without any fine-tuning.
Overworld just dropped Waypoint-1.5, an open-source world model that lets you explore and interact with generative environments in real time. Running locally on everyday GPUs like an RTX 3070 at up to 720p and 60 fps, it effectively shrinks what once required a GPU cluster down to a home gaming rig.
ACE Music just released ACE-Step 1.5 XL, an open-source 4B DiT music generator capable of producing full vocal tracks on a local PC in under a minute. The distilled version runs in just 8 steps and rivals leading commercial tools, though smooth local execution requires at least 12GB of VRAM.
Researchers just open-sourced INSPATIO-WORLD, a real-time simulator that transforms monocular video into an interactive 3D scene you can actually walk through. Instead of guessing raw pixels, it maintains a persistent internal representation, hitting 10 fps on a single consumer RTX 4090. Why it matters: interactive dynamic world models are finally running on everyday GPUs.
Ai2 just open-sourced WildDet3D, a lightweight vision model that estimates an object's real-world 3D position, size, and orientation from a single RGB photo or camera stream. It supports text prompts, point clicks, or 2D boxes, and can even run in real time on an iPhone. Why it matters: It gives robots and AR apps full spatial intelligence without requiring complex multi-camera rigs.
Z.ai just released GLM-5.1, an open-source flagship coding model licensed under MIT that scored 58.4 on SWE-Bench Pro, edging out GPT-5.4 and Claude Opus 4.6. Designed for complex long-horizon tasks, it maintains steady performance gains over hundreds of iterations and thousands of tool calls. Open-source agentic coding is officially competing at the very frontier.
Nous Research just shipped Andrej Karpathy's LLM-Wiki pattern as a native skill inside Hermes Agent. Your agent can now automatically build persistent Markdown knowledge bases that compile, cross-reference, and lint themselves over time. Running a quick update and typing one command basically turns the agent into a tireless research librarian.
Ben Sigman and actress Milla Jovovich just released MemPalace, an open-source AI memory system built with Claude that achieved the first perfect 500/500 score on LongMemEval. Instead of sending data to the cloud, it runs locally, organizing memories into a palace architecture and compressing user context down to ~120 tokens.
SpatialEdit just dropped as a fully open-source 16B model that gives you fine-grained geometric control over 2D images. Beyond shifting objects and rotating their angles, it lets you tweak camera yaw, pitch, and zoom with remarkable precision. The catch? The 32 GB weights mean you'll need a beefy GPU to run it locally.
Nous Research just rolled out Hermes Agent v0.7.0, overhauling its memory architecture into an extensible plugin system. It features built-in memory right out of the box, six third-party providers, and the freedom to bring your own backend via hermes memory setup. Giving agents modular, flexible memory just got much simpler.
