open-source

Microsoft Research released VibeVoice-ASR-BitNet, a quantized 1.58GB version of its VibeVoice-ASR speech recognition model that achieves real-time inference (RTF < 1) on as few as 3 CPU threads. The model runs 1.6-2.3x faster than Whisper.cpp on commodity x86 and ARM hardware, with a modest accuracy tradeoff.

Microsoft has released Mage-Flow, a family of 4B-parameter image generation and editing models built on a shared tokenizer-transformer stack. According to Microsoft, the Turbo variants match or beat open-source systems with 5-8x more parameters while running in 4 diffusion steps.

Within days, Moonshot AI and Alibaba unveiled what they claim are frontier-class models. Moonshot's Kimi K3, at 2.8 trillion parameters, and Alibaba's Qwen3.8, at 2.4 trillion parameters, will both be released as open-weight models with full weights available for download.

LM Studio released Bionic, a Mac app that runs open-source AI models locally or via cloud for coding, document processing, and research tasks. The app includes offline voice transcription using Mistral's Voxtral model and supports models like GLM 5.2 and Kimi K2.7 Code for codebase editing.

Cohere and Cohere Labs released Cohere Transcribe Arabic, a 2B parameter automatic speech recognition model optimized for Arabic dialects and Arabic-English code-switching. The open-source model achieves a 25.9% average word error rate across major Arabic ASR benchmarks, outperforming models up to 30B parameters.

Chinese AI models including DeepSeek and Z.ai have captured over 30% of weekly token usage by U.S. companies on OpenRouter since February 2025, up from 4.5% in the first half of the year. The shift comes as companies seek alternatives 60-90% cheaper than leading models from OpenAI and Anthropic, while Chinese models close the performance gap to within 6-9 months of U.S. frontier systems.

Nex AGI has released Nex-N2-Mini, an open-source agentic mixture-of-experts model with a 262K-token context window. The model accepts text and image inputs and is priced at $0.025 per 1M input tokens and $0.10 per 1M output tokens.

Mistral AI has released Leanstral 1.5, an open-source 119B parameter mixture-of-experts model designed specifically for Lean 4 proof assistance. The model features 128 experts with 4 active per token (6.5B activated parameters), a 256k token context window, and multimodal input capabilities.

Portugal has released Amália, its first national AI model trained specifically for European Portuguese. Built on EuroLLM-9B with 9 billion parameters, the model is fully open-source with weights, datasets, and code published under an open license. The government has committed €5.5m in initial funding through 2027.

DeepReinforce has released Ornith-1.0, an MIT-licensed model designed for agentic coding tasks with variants ranging from 9B to 397B parameters. Built on top of Apache 2.0-licensed Gemma 4 and Qwen 3.5 base models, the company claims it achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks.

DeepSeek has released two new MoE models: DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated). Both models support a one million token context window and use a hybrid attention architecture that requires only 27% of single-token inference FLOPs and 10% of KV cache compared to DeepSeek-V3.2.

The open model ecosystem is diversifying beyond its previous Chinese dominance, with companies like NVIDIA, Cohere, Poolside, and Zyphra releasing models under permissive licenses. NVIDIA's 550B parameter Nemotron-3-Ultra uses LatentMoE architecture and switches to the OpenMDW license, while Cohere released Command A+ as a 218B-A25B MoE under Apache 2.0.

DeepSeek has released two Mixture-of-Experts models: V4-Pro with 1.6 trillion parameters (49B activated) and V4-Flash with 284B parameters (13B activated), both supporting 1 million token context windows. V4-Pro requires only 27% of inference FLOPs and 10% of KV cache compared to V3.2 at 1M token context, trained on over 32 trillion tokens.

Zhipu AI's open-source GLM-5.2 model scores within one percentage point of Anthropic's Claude Opus 4.8 on a key agentic benchmark while costing approximately one-fifth as much. The release comes as U.S. government restrictions limit access to Anthropic's Fable and OpenAI's GPT-5.6 models.

Z.ai's GLM-5.2 performs on par with Claude Opus 4.8 and OpenAI's GPT-5.5 in cybersecurity benchmarks while costing roughly half as much to run. Security evaluations from Graphistry and Semgrep confirm the open-weight model's capabilities in vulnerability discovery and cyber investigation, raising concerns about accessibility of advanced hacking tools.

Trail of Bits launched Patch the Planet, a security initiative using OpenAI's GPT-5.5-Cyber model to find and fix bugs in critical open-source projects. The first week produced 64 pull requests and 51 issues across 19 projects including cURL, Python, PyPI, and Sigstore, with 37 patches already merged.

Baidu has released Unlimited-OCR, a 3 billion parameter model for optical character recognition and document parsing. The model supports single-page and multi-page document processing with a 32,768 token context window and runs on NVIDIA GPUs using bfloat16 precision.

Poolside has released Laguna M.1, a 225B total parameter Mixture-of-Experts model with 23B activated parameters per token, designed for agentic coding tasks. The model scores 74.6% on SWE-bench Verified and 63.1% on SWE-bench Multilingual, released under Apache 2.0 license.

Mistral AI has released Leanstral, a sparse 120B model with 6B active parameters designed specifically for the Lean 4 proof assistant. The model is available under Apache 2.0 license with free API access and achieves a 26.3 FLTEval score at pass@2, outperforming Claude Sonnet 4.6 while costing $36 versus $549.

Mistral has released Mistral 3, including Mistral Large 3—a sparse mixture-of-experts model with 41B active and 675B total parameters—and three Ministral 3 edge models (3B, 8B, 14B). All models are released under Apache 2.0 license with multimodal capabilities and are available today on multiple platforms.

Zhipu AI released GLM-5.2, a 753 billion parameter model with a 1 million token context window. The model scores 62.1% on SWE-bench Pro and introduces IndexShare architecture that reduces per-token FLOPs by 2.9× at 1M context length. Released under MIT license with no regional restrictions.

Z.AI has released GLM-5.2, an open-source model with a 1M-token context window under an MIT license. On FrontierSWE, a long-horizon coding benchmark, GLM-5.2 trails Claude Opus 4.8 by 1% while outperforming GPT-5.5 by 1%, and achieves 81.0 on Terminal-Bench 2.1 compared to Opus 4.8's 85.0.

Zhipu AI has released GLM-5.2, a 753 billion parameter model featuring a 1 million token context window and MIT open-source license. The model scores 62.1% on SWE-bench Pro and 91.2% on GPQA-Diamond, with flexible reasoning effort levels for coding tasks.

Microsoft released FastContext-1.0, a lightweight repository-exploration subagent for LLM coding agents spanning 4B to 30B parameters. The model reduced main-agent token consumption by up to 60% while improving end-to-end resolution rates by up to 5.5% on SWE-bench Pro when integrated with agents like GPT-5.4 and GLM-5.1.

Google DeepMind has released Gemma 4 12B, a 12-billion parameter multimodal model that runs locally on laptops with 16GB of RAM. The model eliminates separate vision and audio encoders, processing raw inputs directly through its language model backbone under an Apache 2.0 license.

Amazon has released an open-source testing framework for Nova Sonic voice agents that automates multi-turn conversation evaluation without requiring human testers. The harness uses LLM-as-judge techniques to assess voice agents across six metrics including goal achievement, response accuracy, and tool usage, addressing a critical QA bottleneck in voice AI development.

Google DeepMind has released Gemma 4 12B Unified, an encoder-free multimodal model that processes text, images, and audio through a single decoder-only transformer. The model features 11.95 billion parameters, a 256K token context window, and achieves 77.2% on MMLU Pro and 72.0% on LiveCodeBench v6.

ByteDance released Bernini-R, an open-source video generation and editing model that combines an MLLM-based semantic planner with a DiT-based renderer. The model requires Hopper-class GPUs (H100/H800/H200) for optimal performance and supports multiple tasks including text-to-video, video editing, and reference-guided generation.

JetBrains has released Mellum2-12B-A2.5B-Thinking, a reasoning-augmented assistant model with 131,072-token context window and 64 Mixture-of-Experts architecture that activates 8 experts per token. The model emits explicit chain-of-thought reasoning inside <think> blocks before providing final answers.

NVIDIA has released Cosmos 3, a unified omni-model that combines world generation, physical reasoning, and action generation in a single architecture. Available in 8B (Nano) and 32B (Super) parameter versions on Hugging Face, Cosmos 3 uses a Mixture-of-Transformers architecture to process text, image, video, audio, and action modalities without switching between separate models.

Mistral AI has released Mistral Small 4, a 119B total parameter mixture-of-experts model with 256K context window and native multimodal capabilities. The model uses 128 experts with 4 active per token (6B active parameters) and is released under the Apache 2.0 license, marking Mistral's first unified model combining reasoning, multimodal, and coding capabilities.

Mistral AI released Leanstral, the first open-source code agent designed specifically for Lean 4 formal proof verification. The model uses 6B active parameters in a sparse 120B architecture and is available under Apache 2.0 license with free API access.

Mistral AI has released Mistral 3, consisting of Mistral Large 3—a sparse mixture-of-experts model with 675B total parameters and 41B active parameters—and three Ministral 3 models at 3B, 8B, and 14B parameters. All models are released under the Apache 2.0 license with multimodal capabilities including image understanding.

Mistral AI released Voxtral, a family of open-source speech understanding models available in 24B and 3B parameter variants under Apache 2.0 license. The models support up to 32K token context (30 minutes of audio for transcription, 40 minutes for understanding) and are priced at $0.001 per minute via API—less than half the cost of comparable proprietary systems according to Mistral.

Mistral AI has released two specialized coding models: Devstral Medium, achieving 61.6% on SWE-Bench Verified, and Devstral Small 1.1, scoring 53.6% and released under Apache 2.0 license. The company claims Devstral Medium surpasses Gemini 2.5 Pro and GPT-4.1 at a quarter of the price.

Mistral AI has released Magistral, its first reasoning model line, in two versions: Magistral Small (24B parameters, Apache 2.0) and Magistral Medium (enterprise). Magistral Medium scored 73.6% on AIME2024 (90% with majority voting at 64 samples), while the open-source Small version achieved 70.7% (83.3% with voting).

Tencent released Hy-MT2, a family of multilingual translation models available in 1.8B, 7B, and 30B-A3B (MoE) sizes. All models support translation among 33 languages and follow translation instructions in multiple languages. The 1.8B model can be compressed to 440MB using 1.25-bit AngelSlim quantization.

NVIDIA has released Nemotron-Labs Diffusion, a family of diffusion language models at 3B, 8B, and 14B scales that generate multiple tokens in parallel rather than one at a time. The 8B model achieves 6.4× higher tokens per forward pass than autoregressive models in self-speculation mode while maintaining comparable accuracy.

Tencent has open-sourced Hy-MT2, a family of multilingual translation models available in 1.8B, 7B, and 30B-A3B parameter sizes. The models support translation across 33 languages and include extreme quantization down to 1.25-bit, reducing the 1.8B model to 440MB storage while increasing inference speed by 1.5x.

Cohere has released Command A+ as an open source model under Apache 2.0 license. The sparse mixture-of-experts architecture features 25 billion active parameters out of 218B total parameters, supports 128K input context length, and includes vision capabilities alongside tool use and reasoning features.

Cohere has released Command A+, an open-source sparse mixture-of-experts model with 218 billion total parameters and 25 billion active parameters. The model features W4A4 quantization allowing deployment on a single Nvidia B200 GPU, supports 128K input context, and includes built-in chain-of-thought reasoning with vision capabilities.

IBM Research has launched the Open Agent Leaderboard, the first open benchmark that evaluates complete AI agent systems rather than just underlying models. The leaderboard reveals that agents using identical models can achieve significantly different success rates and costs depending on system architecture, with failed runs costing 20-54% more than successful ones.

IBM released two new multilingual embedding models under Apache 2.0: a 97M-parameter compact model scoring 60.3 on MTEB Multilingual Retrieval (highest in its size class) and a 311M full-size model scoring 65.2. Both support 200+ languages with enhanced retrieval for 52 languages, handle 32K-token context (64x increase over predecessors), and include code retrieval across 9 programming languages.

IBM released Granite Embedding 311M Multilingual R2, a 311-million parameter dense embedding model with 32,768-token context length and support for 200+ languages. The model scores 64.0 on Multilingual MTEB Retrieval (18 tasks), an 11.8-point improvement over its predecessor, and ships with ONNX and OpenVINO models for production deployment.

IBM has released the Granite 4.1 family of language models under Apache 2.0 license. The models come in 3B, 8B, and 30B parameter sizes. Unsloth has released 21 GGUF quantized variants of the 3B model ranging from 1.2GB to 6.34GB.

IBM released Granite 4.1 30B, a 30-billion parameter instruction-following model with a 131,072 token context window. The model scores 80.16 on MMLU 5-shot and 88.41 on HumanEval pass@1, with enhanced tool-calling capabilities following OpenAI's function definition schema.

IBM has released Granite 4.1 8B, an 8-billion-parameter decoder-only language model with a 131,072-token context window. The model supports 12 languages and costs $0.05 per million input tokens and $0.10 per million output tokens, available under the Apache 2.0 license.

IBM has released Granite 4.1-8B, an 8-billion parameter long-context model with a 131,072-token context window. The model achieves 85.37% on HumanEval and 73.84% on MMLU 5-shot, with enhanced tool-calling capabilities reaching 68.27% on BFCL v3. Released under Apache 2.0 license, it supports 12 languages.

Poolside has released Laguna XS.2, a 33B total parameter Mixture-of-Experts model with 3B activated parameters per token, designed for agentic coding. The model features a 131,072-token context window, scores 68.2% on SWE-bench Verified, and is available under Apache 2.0 license with free API access.

Xiaomi released MiMo-V2.5, a 310B total parameter sparse mixture-of-experts model that activates 15B parameters per token. The omnimodal model supports text, image, video, and audio understanding with a 1M token context window and was trained on 48T tokens using FP8 mixed precision.