New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
xAI just open-sourced Grok 2.5, releasing the weights for its top 2024 model on Hugging Face under a custom community license. It is free for research and non-commercial projects but bans training rival models. Musk also promised Grok 3 will go open source in about six months.
Elon Musk just announced that xAI will open-source its Grok 2 chatbot next week, sticking to his promise of releasing previous-generation models as new ones arrive. While giants like OpenAI lean into proprietary APIs, xAI is betting that handing developers the actual code will carve out a serious competitive edge.
Researchers at The University of Hong Kong just launched OpenCUA, an open source framework designed to train AI agents that control computers across Windows, macOS, and Ubuntu. Powered by 22,600+ screen-recorded task demonstrations and chain-of-thought reasoning, it matches proprietary models from Anthropic and OpenAI. Why it matters: It finally gives the research community an open recipe to study and build general-purpose desktop agents.
ByteDance's Seed Team just open-sourced Seed-OSS-36B on Hugging Face under the Apache-2.0 license. The 36B-parameter model boasts a 512K token context window—doubling GPT-5's limit—and lets developers customize "thinking budgets" to balance reasoning quality with latency. It gives enterprises a fully permissive, commercially viable alternative for heavy coding and long-document reasoning.
Nvidia just dropped Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer model slimmed down to fit on a single A10 GPU. Developers can toggle reasoning traces on or off using simple /think tokens and dial in exact thinking budgets. It delivers up to 6x faster inference than similar transformers while outperforming Qwen3-8B across key benchmarks.
NVIDIA just released Granary, an open-source multilingual speech dataset spanning around a million hours of audio across 25 European languages. Alongside it, the company launched Canary-1b-v2 for high-accuracy translation and Parakeet-tdt-0.6b-v3 for fast, real-time transcription on Hugging Face.
Google just released Gemma 3 270M, an ultra-compact 270-million-parameter open model built specifically for task-specific fine-tuning. Packed with out-of-the-box instruction following and structuring capabilities, it lets developers ditch bloated models in favor of cheap, fast specialized workers for jobs like text classification and data extraction.
Roblox just open-sourced Sentinel, an AI safety system built to detect grooming and child endangerment. The tool scans over six billion chat messages daily in real time, using one-minute snapshots to catch coded and indirect language. In early 2025, it proactively caught 35% of detected cases before any user report was filed.
Skywork just launched Matrix-Game 2.0, introduced as the first open-source, real-time, long-sequence interactive world model. While DeepMind kept last week's Genie 3 behind closed doors, Skywork is making its model fully open-source, delivering 25FPS rendering and minutes-long interactions.
Microsoft just rolled out GPU-optimized variants of OpenAI's gpt-oss-20B model for Windows devices, enabling local reasoning on the edge. Windows developers can fire it up in minutes using Foundry Local or the AI Toolkit for VS Code to test prompts and hook it directly into local workflows.
OpenAI just released gpt-oss-120b and gpt-oss-20b, its new open-weight reasoning models under the Apache 2.0 license. Both support full chain-of-thought and tool use, with the 20b model running locally on just 16 GB of memory. Why it matters: developers can now run top-tier reasoning systems entirely on their own consumer hardware.
Alibaba just released Qwen-Image, an open-source 20B MMDiT foundation model built to tackle complex text rendering in both English and Chinese. It handles multi-line typography, paragraph layouts, and precision editing without garbling letters. It delivers native bilingual graphic design directly to the open-source image ecosystem.
Krea just released open weights for FLUX.1 Krea [dev], an image model built alongside Black Forest Labs designed to ditch the generic "AI look." Fully compatible with the FLUX.1-dev pipeline, it relies on opinionated post-training with fewer than 1M hand-curated samples, ditching waxy textures and artificial blur for authentic, photorealistic aesthetic control.
Tencent Hunyuan just open-sourced Hunyuan3D World Model 1.0, an industry-first model that turns a single prompt or image into an explorable, interactive 3D scene. It connects directly with standard CG pipelines for editing and simulation, built to speed up game dev and VR creation.
Mistral just released Voxtral, open-sourcing two state-of-the-art speech understanding models under Apache 2.0 in 24B and 3B sizes. With a 32k context window handling up to 40 minutes of audio, it outperforms Whisper large-v3, supports direct function calling straight from voice, and offers an API at less than half the price of competitors.
OpenAI is hitting pause on its open-weight model, which was scheduled to drop next week. CEO Sam Altman said the team needs more time to run extra safety tests and review high-risk areas. Why it matters: once model weights are released, they can never be recalled, so OpenAI is choosing caution over speed.
Moonshot AI just open-sourced Kimi K2, a 1-trillion parameter MoE model with 32B activated parameters engineered specifically for coding and autonomous agent workflows. It scored 53.7% on LiveCodeBench, comfortably beating DeepSeek-V3 and GPT-4.1, powered by a new MuonClip optimizer that eliminates training instability. Why it matters: High-end agentic intelligence is rapidly shifting from pricey closed APIs to open, dirt-cheap alternatives.
Reka just open-sourced Flash 3.1, a 21B model with major coding upgrades, alongside its new Reka Quant library. The team managed to squeeze the model down to 3.5 bits for llama.cpp with minimal accuracy loss, giving developers a compact agent base that punches well above its weight.
Mistral AI just teamed up with All Hands AI to roll out the Devstral models built specifically for coding agents. The 24B Devstral Small 1.1 is fully open under Apache 2.0 and hits 53.6% on SWE-Bench Verified, while Devstral Medium reaches 61.6% on the benchmark via API at roughly a quarter of the price of frontier competitors.
Google just released MedGemma 27B Multimodal and a 400M-parameter image encoder called MedSigLIP for health AI development. Both run on a single GPU—with lightweight variants even fitting on mobile devices—to help researchers build private, tailored medical tools without relying on closed APIs.
