open source

Microsoft Research has published Probabilistic Bias Correction (PBC) in Nature and released it as open-source software. The researchers report it improved a leading ECMWF physics-based model for 98% of precipitation targets and 92% of temperature targets at 2–6 week lead times, and that it runs on CPUs.

Anthropic has launched OSS Scanner, an opt-in service that gives open-source projects periodic vulnerability scans at no cost, run by its strongest models including Claude Mythos. Reports are fully model-generated with no human review or triage, so Anthropic warns some may be incorrect or invalid.

Meta introduced Muse Gadgets on Friday, an open-source project with firmware and a Linux SDK for building hardware that connects to its Muse personal AI agent. Meta also built a first-party device, Muse Home Link, and says it is giving away 5,000 units to Muse subscribers.

Amazon Web Services has released Strands Decider 2B, an open-source model that chooses among pre-decided options and returns a confidence score instead of generating text. It is inspired by TypeSafe's Jev and is small enough to run locally. Amazon says it briefly topped the Jevbench ranking for models of its size.

Ai2 released Olmo-core 3, an open training framework for large mixture-of-experts models. The company reports about 2.7× the throughput of its earlier FSDP-based implementation and benchmarks up to 1.2 trillion total parameters on 512 NVIDIA B300 GPUs. It is the infrastructure for Ai2's next MoE-based Olmo, not a model release.

Black Forest Labs has released FLUX 3 Action, an open-weight robotics model built on its FLUX 3 multimodal foundation. The 7-billion-parameter model reads multi-camera video feeds and predicts what a robot should do next, claiming a record success rate on the RoboLab-120 leaderboard while running nearly 4x faster than the previous best open model.

China Telecom Artificial Intelligence Technology has released Xing4.0-29B-A4B, a 29-billion-parameter mixture-of-experts model with only 4B parameters active per token and native 256K context. The company claims it is the first model of this scale trained entirely on Huawei's Ascend NPU platform using the MindSpore framework.

Mistral AI will power Mozilla's Firefox Smart Window (beta), a browsing assistant that summarizes searches and tracks tabs. The rollout starts in France and North America, with the UK and Germany to follow later this year.

IBM has released Granite Time Series PatchTST-FM-r2, a 385M-parameter forecasting model that ranks #2 overall among replicable zero-shot models on the GIFT-Eval benchmark and first among those with permissive, commercial-friendly licensing. The model uses conformer-based architecture and supports probabilistic forecasts across up to 8,192-step contexts.

Alibaba has released Qwen-Drive 1.0, a driving model built on Qwen3.5-4B that handles spatial perception, route planning, and cockpit dialogue in a single system. Reinforcement learning cut the rate of off-road driving errors in simulation from 24 percent to 12 percent, though the model's stated reasoning doesn't always match its actual maneuvers.

Microsoft Research has released VibeVoice-ASR-Streaming-7B, an open-weight streaming automatic speech recognition model that transcribes both who is speaking and what they say in real time. The model, listed at 9B parameters despite its name, supports 10 languages and custom hotwords under an MIT license.

Tencent has released WeMM-Embedding-9B, a 9-billion-parameter multimodal embedding model built on Qwen3.5 that produces 4,096-dimensional embeddings from text, images, video, and visual documents. The model reports state-of-the-art results on the MMEB-v2 and MMEB-v3 benchmarks and is released under Apache 2.0.

Hugging Face's transformers v5.16.1 release adds support for GLM-5.3-Flash, a 320B total-parameter (18B active) multimodal mixture-of-experts model. Zhipu AI claims it outperforms GLM-5.2 while approaching Claude Opus 4.8 on coding and agentic benchmarks at one-tenth the cost.

IBM has released two new 470M-parameter Granite Speech models that transcribe audio at over 12,600x real-time speed on an NVIDIA H200 GPU. The encoder-only models achieve 4.85-5.00% word error rate while running over 20x faster than prior Granite Speech versions.

Meta has released Muse Glimmer, a 30B parameter multimodal model distilled from its larger Muse model, designed for local, privacy-aware agentic applications. The model ships under the Apache 2.0 license as part of Hugging Face Transformers v5.15.0, which also adds support for IBM's GraniteSWA/GraniteMoeSWA and the A.X-K1/K2 models.

NVIDIA has released NemotronLabs VoiceChat 11B, an 11-billion-parameter end-to-end full-duplex speech model that unifies streaming speech understanding and generation in one architecture. The model claims to be the first open full-duplex system to support live tool calling during natural conversation, with ~450ms turn-taking latency.

AMD has released Instella-MoE-16B-A3B-Think, a 16-billion-parameter Mixture-of-Experts language model trained entirely from scratch on AMD Instinct MI300X and MI325X GPUs. The release includes every checkpoint from pre-training through reinforcement learning, along with full training recipes, under a research-only license.

Moonshot AI has released open weights, a technical report, and supporting infrastructure for Kimi K3, a model that claims 2.5x more intelligence per unit of compute. Independent testing found notable gaps in cybersecurity and math performance compared to Western frontier models.

Chinese company Moonshot AI released Kimi K3, an open source model that the company claims demonstrates frontier-level performance while trailing only Claude Fable 5 and GPT 5.6 Sol. Independent analyses from Arena.ai and Vals AI suggest the model is competitive with flagship frontier models, reigniting debate about Chinese AI capabilities and open source model development.

Mistral AI released Leanstral 1.5, a free Apache-2.0 licensed model with 119B total parameters and 6B active parameters specialized for formal verification in Lean 4. The model achieves 100% on miniF2F benchmark, solves 587 of 672 PutnamBench problems at $4 per problem (versus $300+ for competitors), and reaches state-of-the-art 87% on FATE-H and 34% on FATE-X benchmarks.

Microsoft is evaluating a self-hosted version of DeepSeek V3 to power Copilot Cowork as agent costs spiral. The company plans to launch a lower-cost tier within weeks while moving to usage-based pricing, charging enterprises for actual compute consumed rather than flat fees.

Cohere Labs has released North Mini Code 1.0, a 30B-parameter sparse Mixture-of-Experts model with 3B active parameters and a 256K context window. The Apache 2.0-licensed model is optimized for agentic software engineering, featuring 128 experts with 8 activated per token, and trained specifically for tool use in coding tasks.

NVIDIA has released Nemotron 3.5 Content Safety, a 4B-parameter model built on Google Gemma 3 4B IT that provides multimodal safety classification across approximately 140 languages. The model includes a 128K context window, custom enterprise policy enforcement, auditable reasoning traces, and is releasing its training dataset.

JetBrains has released Mellum2, a 12-billion parameter Mixture-of-Experts model that activates only 2.5 billion parameters per token. The open-source model is designed for code generation, RAG pipelines, and agent workflows with 2x faster inference than similar-sized models.

Moonshot AI has released Kimi K2.6, an open-source multimodal model with 1 trillion total parameters (32B activated) and 256K context window. The model achieves 80.2% on SWE-Bench Verified, 58.6% on SWE-Bench Pro, and supports horizontal scaling to 300 sub-agents executing 4,000 coordinated steps.