New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Google DeepMind just open-sourced Gemma Scope 2, an interpretability suite designed to reverse-engineer how the Gemma 3 family actually thinks. Spanning models from 270M to 27B parameters, it uses sparse autoencoders and transcoders across every layer to map internal computations. Why it matters: It gives the safety community an open, microscopic look into why models hallucinate, get jailbroken, or disguise their reasoning.
Anthropic just released Bloom, an open-source tool that automatically generates targeted evaluation suites to detect misaligned AI behaviors. Instead of researchers manually building static benchmarks that risk training contamination, Bloom takes a target trait—such as sycophancy or sabotage—and dynamically generates multi-turn scenarios to measure its frequency. Why it matters: It cuts evaluation development from weeks to days, helping safety researchers keep up with rapid frontier model updates.
LMArena just open-sourced Arena-Rank, the standalone Python package behind its widely watched AI model leaderboards. Rewritten with a JAX backend, the toolkit calculates rankings over 30x faster than before and introduces reweighting to fairly evaluate models with fewer match records. Beyond LLM evaluations, it can also rank head-to-head match data in sports and gaming.
Microsoft researchers just released TRELLIS.2-4B, an open-source 4-billion-parameter model that converts a single image into high-fidelity 3D assets with full PBR materials. Built on a novel sparse voxel structure called O-Voxel, it can produce 512³ resolution assets in about 3 seconds on an H100 GPU, though it demands an NVIDIA GPU with at least 24GB of memory.
Xiaomi just open-sourced MiMo-V2-Flash, a 309B MoE foundation model with 15B active parameters tailored for coding and reasoning. It ranks #1 among open-source models on SWE-bench Verified, pumps out 150 tokens per second, and costs just $0.1 per million input tokens. Why it matters: High-performance agentic coding is becoming dramatically faster and cheaper for everyday builders.
NVIDIA just debuted Nemotron 3, a family of open models built specifically for multi-agent AI systems. Leading the pack is the lightweight Nano model, which quadruples token throughput over its predecessor using a hybrid MoE design, backed by open-source datasets and RL libraries.
AI2 just released Molmo 2, an open-source multimodal model family built for video pointing, tracking, and question answering. Packing 4B, 7B, and 8B variants alongside a full training and inference codebase, the 8B model punches way above its weight—topping Gemini 3 Pro on video tracking while training on under one-eighth the video data of Meta's PerceptionLM. Why it matters: developers now get frontier-grade spatio-temporal video grounding in an efficient, completely inspectable open stack.
DiffSynth-Studio just released Qwen-Image-i2L, an open-source suite that transforms a single image directly into a custom LoRA model. Built on SigLIP2, DINOv3, and Qwen-VL, it provides variants ranging from a 2.4B aesthetic extractor to a 7.9B coarse content model. Why it matters: Creators no longer need large datasets or tedious fine-tuning runs just to capture a unique style or subject.
Mistral AI just released Devstral 2, a lightweight open-weights coding model family led by a 123B flagship and a 24B small version. The larger model scores 72.2% on SWE-bench Verified at up to 7x the cost efficiency of Claude Sonnet, while the 24B version runs locally on consumer hardware. The release also includes Mistral Vibe, an open-source terminal CLI for autonomous software development.
Anthropic just donated the Model Context Protocol (MCP) to the newly formed Agentic AI Foundation under the Linux Foundation. Co-founded alongside Block and OpenAI, the move makes the runaway agent connector standard officially vendor-neutral. With over 10,000 active public servers and native support in ChatGPT, Cursor, and Gemini, MCP is cementing its role as the web's default plumbing for AI agents.
Zhipu AI just open-sourced the GLM-4.6V series, offering a 106B flagship and a lightweight 9B Flash model under the MIT license. The standout upgrade is native multimodal tool calling, which lets agents feed raw images and screenshots straight into external tools without losing detail across a 128k context window. It gives developers an enterprise-ready open alternative.
Mistral AI just unleashed Mistral 3, an entire family of multimodal models released under a fully permissive Apache 2.0 license. The drop includes Mistral Large 3, a massive 675B-parameter MoE model (41B active), alongside edge-friendly Ministral 3 variants spanning 3B, 8B, and 14B sizes with native vision support. Why it matters: developers get frontier-grade open weights they can actually modify and deploy commercially without restrictive licensing handcuffs.
Hugging Face just rolled out Agent Skills, standardized instruction packages that teach coding agents like Claude Code, Cursor, and Codex how to train models, manage datasets, and run evals. Each capability loads on demand through a simple SKILL.md file, basically turning generic coding assistants into full-stack ML engineers across the HF ecosystem.
NVIDIA just expanded its open-source footprint at NeurIPS with Alpamayo-R1, the first industry-scale open reasoning VLA model for autonomous driving. The model pairs chain-of-thought reasoning with path planning, helping vehicles handle messy edge cases like crowded intersections by breaking down situations step by step. Why it matters: bringing common-sense reasoning into physical AI could accelerate the path toward safe level 4 autonomy.
Microsoft just open-sourced Fara-7B, a compact 7B-parameter model specifically built to take over keyboard and mouse actions. By reading raw browser screenshots instead of code trees, it automates routine web tasks like booking trips or comparing prices. Why it matters: It is small enough to run locally on Copilot+ PCs, delivering fast agentic workflows without uploading private screen data.
Hugging Face just launched AnyLanguageModel, an open-source Swift package that serves as a drop-in replacement for Apple's Foundation Models framework. By swapping a single import statement, developers can alternate between local engines like MLX and Core ML and cloud providers without rewriting glue code. It dramatically lowers the bar for running open-source models on Apple silicon.
Meta just expanded its computer vision lineup with SAM 3 and SAM 3D. SAM 3 lets you track and segment objects in images and video using natural text prompts, while SAM 3D reconstructs 2D photos into workable 3D objects and human avatars. Both models are available to test in a new interactive playground, with SAM 3 code and weights fully open-sourced.
NVIDIA just unveiled Apollo, a family of open AI physics models designed to speed up scientific and industrial simulation. The models cover domains like semiconductors, fluid dynamics, and climate modeling, using architectures like neural operators and transformers. Industry giants like Siemens, Synopsys, and Cadence are already integrating them to run real-time simulations in seconds.
Microsoft is doubling down on its flagship Earth-system AI model, Aurora. In partnership with the University of Cambridge, the company pledged to keep future releases and full training pipelines completely open-source, aiming to help local weather services and energy grids build tailored forecasting tools with minimal data.
Microsoft Research just introduced MMCTAgent, an open-source framework built on AutoGen to reason across long videos and massive image libraries. Instead of relying on one-shot answers, it deploys Planner and Critic agents to break down queries, trigger computer vision tools, and cross-check evidence. On the MM-Vet benchmark, it pushed GPT-4V's accuracy from 60.20% to 74.24%.
