open weights

Cloudflare has released Clairic, a model that extracts named fields from unstructured text in a single pass instead of generating tokens one at a time. It is in preview on Workers AI, with open weights promised soon. Cloudflare claims a multi-step agent chain using Clairic runs in 19.3 seconds, versus 61.7 seconds for GLM-5.3 alone.

MiniMax has released M3.1 Flash, a reasoning model that scores 42 on the Artificial Analysis Intelligence Index, up from 29 for M3. It costs $0.15 per 1M input tokens and $0.60 per 1M output tokens, and open weights are planned for October 23.

Mistral AI has released Timestral-209M, the largest of four time series foundation models (5M, 28M, 123M, 209M parameters). According to the model card, it leads GIFT-Eval, TIME and fev-bench in zero-shot forecasting, with a context of up to 32,768 time steps and 21-quantile probabilistic output.

Tencent Hunyuan released Hy-Memory-30B-A3B, a 30B-parameter mixture-of-experts model with about 3B active parameters, built to run the write path in external memory frameworks. Tencent reports a 55.49 cross-harness average, ahead of Claude Opus 5 (54.62) and DeepSeek V4.1 Flash (53.64) on its own fixed 20% benchmark subsamples.

FLUX 3 Image Klein is a 5.6B-parameter open-weights image model covering text-to-image, multi-reference generation and image editing. Its developer claims it leads its own blind evaluations in all three tasks. Pricing, license terms and benchmark data were not disclosed in the announcement.

Liquid AI has published GGUF builds of LFM2.5-24B-A2B, a 24B-parameter model on the lfm2moe architecture, in sizes from 13.6 GB (Q4_0) to 48 GB (BF16). The Q4_0 file is a Quantization-Aware Distillation checkpoint trained to recover accuracy lost to 4-bit quantization.

Fish Audio has released Drama 3, a text-to-speech model with 35B total parameters and 3B active per token, built for multi-character, long-form audio. It supports up to 16 speakers per text, voice cloning, and streaming. Fish Audio reports first audio in under 100 ms on a single H200.

OpenBMB has released MiniCPM-V 4.7 1B, a 1B-parameter multimodal model for single-image, multi-image and video understanding on phones, tablets and local hardware. OpenBMB claims it substantially outperforms Qwen3.5 0.8B in overall score and inference efficiency in Instruct mode. Benchmark figures, context window and training cutoff are not given in the available model card text.

Microsoft has released FrogNano-4B-2609, a repository-level coding agent derived from Qwen3.5-4B and published under Apache 2.0 with open weights. Microsoft says it was post-trained only with reinforcement learning on about 1,500 synthetic software-engineering tasks, with no stronger-model trajectories. It is evaluated at roughly 131K tokens of context.

Liquid AI released two open-weight decision models, d1-3B (text and image) and the experimental d1-omni-600M (text with image or audio). Unlike generative models, they answer in a single forward pass, and Liquid AI claims d1-3B scores 48.57 on its Decision Index 0.2.1, ahead of all 4B and 9B models it tested.

Musubi released PolicyLM-1.7B on Tuesday, a 1.7-billion-parameter open-weights decision model built for real-time content moderation. The company claims it applies plain-English policies to messages in under 50 milliseconds and needs no retraining when policies change.

Google announced EmbeddingGemma 2, a 740M-parameter natively multimodal embedding model built on the Gemma 4 architecture and released under Apache 2.0. Google says the quantized model needs about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model on a Pixel 11 Pro. Google also launched a Mac app, AI Edge Foresight, to demonstrate it.

Ideogram has released Ideogram 4.5, an image model it claims edits only the area a user specifies and leaves the rest untouched. It offers four quality tiers from 0.8 to 22 cents per image, all at native 2K resolution, via the Ideogram platform and API. An open-weight release is promised but not yet dated.

Apple has released LensVLM-9B, a 9-billion-parameter vision-language model fine-tuned from Qwen3.5-9B-Base that processes documents as compressed images, selectively expanding only relevant pages to full resolution. The model supports 5x, 10x, and 15x compression ratios and is available under Apple's Machine Learning Research Model License.

Xiaomi released MiMo-V2.6-Pro, a 1.02T-parameter mixture-of-experts model with 42B active parameters, which debuted as the top-scoring open-weights model on Artificial Analysis' Intelligence Index (46). The company claims the model's RL training run cost roughly $2.6M and completed in 130 hours.

Xiaomi's MiMo team has released MiMo-V2.6-Pro-RL, a 1.02-trillion-parameter sparse mixture-of-experts model with 42B active parameters, 1M-token context, and native text/image/video/audio processing. The model was trained via a single mixed reinforcement learning run spanning coding, agentic, visual, and cybersecurity tasks, with benchmark scores that Xiaomi claims approach or match Claude Opus 5 and GPT-5.6 on several agentic and coding tests.

Cactus Compute has released Needle 3, a foundation model compressed to 8-29 MB that runs entirely on-device for tool calling, structured data extraction, and text embedding. The company claims it beats models 10x its size on mobile tool calls while running on hardware as small as microcontrollers.

Shanghai Artificial Intelligence Laboratory has released Atria Dawn Preview, a text-only agentic model built on the 744B-parameter MoE GLM-5.2 foundation model with a 256K context window. The model targets multi-step research, coding, and productivity tasks, with benchmark results claimed to compete with DeepSeek V4 Pro, Kimi K3, and Claude Opus 5.

DeepSeek has released V4.1-Flash, a 552B-parameter multimodal Mixture-of-Experts model supporting 1M-token context and activating only 8B parameters during prefill. The model uses a new Causal Encoder-Decoder architecture and Compressed Sparse Attention 2 to cut global KV cache to 890 bytes per token, roughly a quarter of its predecessor.

Alibaba's Qwen team released Qwen3.8-2.4T-A95B on August 12, 2026, the open-weight version of Qwen3.8-Max and the first Qwen-Max-class model made publicly available. The 2.4 trillion-parameter mixture-of-experts model activates only 95 billion parameters per token and supports context windows up to 1 million tokens.

Meta Superintelligence Labs shipped Muse Spark 1.3, which the company claims ranks #3 globally on the Artificial Analysis Intelligence Index and matches OpenAI's GPT-5.6-Sol on coding and agentic benchmarks. The model is available now via Muse Code and Meta's API, with open weights and a follow-up model promised soon.

IBM has released Granite 4.2 8B, a dense reasoning model built for math, code generation, and agentic workflows. The model supports 131K context, 12 languages, and three switchable reasoning modes, priced at $0.10 per 1M input tokens and $0.15 per 1M output tokens.

Hugging Face repositories for Qwen3.8-27B, a matching FP8 quantized build, and LiquidAI's LFM2.5-VL-3B surfaced within the same news cycle. Neither Alibaba's Qwen team nor LiquidAI has published accompanying benchmarks, pricing, or technical reports as of this writing.

MiniMax released Music 3, an open-weight music generation model that produces complete songs up to five minutes long from lyrics and text descriptions. The model combines an 8B and 0.6B language model pair with a Flow Matching synthesis system to output 32 kHz stereo audio.

Hugging Face repositories for Qwen3.8-2.4T-A95B, its FP8 variant, and NVIDIA's Nemotron-3.5-Lightning-30B-A3B have surfaced, but neither company has published accompanying benchmarks, technical reports, or pricing.

Liquid AI has released LFM2.5-VL-3B, a 3.1-billion-parameter vision-language model built for on-device deployment. The model claims leading performance in its size class on grounding, screen understanding, and tool use, while running at 228 tokens/s on an Apple M5 Max.

Lightricks has released LTX-2.5, an open-weight world model that generates synchronized video and audio from text, image, and video inputs. The 22B-parameter model adds native multishot generation, a new diffusion video decoder, and a custom Gemma4 12B text encoder.

NVIDIA has released Alpamayo 2 Super, a 34B-parameter foundation model for autonomous vehicle development that combines a 32B vision-language backbone with a 2.3B-parameter diffusion action decoder. The model handles trajectory prediction, visual question answering, 2D grounding, and auto-labeling, and posts a Lingo-Judge score of 79.2 on LingoQA reasoning evaluation.

NVIDIA's Magpie TTS Multilingual, a 364M-parameter open-weights text-to-speech model, now supports 12 languages after adding Modern Standard Arabic, Korean, and Brazilian Portuguese. The model achieves 32ms time-to-first-audio on B200 GPUs and improves speech quality across French, Spanish, and German.

Unsloth has published GGUF quantizations of Muse Glimmer-30B, a dense 29.6B-parameter causal transformer with a dedicated perception encoder, attributed to Meta Superintelligence Lab in the model card. The model targets autonomous agentic tasks on consumer hardware with a 131,072-token context window and 4-bit quantization under 20GB.

Liquid AI has released LFM2.5-2.6B, a 2.6B-parameter model designed to run capable tool-calling agents locally on laptops and phones. The company claims it matches or beats models up to 4x its size on instruction-following and tool-use benchmarks while running under 2.5GB of memory.

Mistral has released Shieldstral, a 3B open-weights safety classifier that reframes content moderation as a policy-adaptive question-answering task. The model claims to match or outperform guard models up to 7x its size on text safety and multimodal benchmarks, and runs on a single 16GB GPU.

MiniMax has released open weights for H3, a 33-billion-parameter video model that ranks first in Video Editing and second in Text-to-Video on Artificial Analysis — the first time an open model has topped a video generation category. The model accepts text, images, video, and audio in a single prompt, though its highest-resolution module remains closed.

A wave of open-weight model releases from Thinking Machines, Tencent, Poolside, Moonshot AI, and Meituan signals that model-building capacity is spreading rather than consolidating. The releases range from a 1.6 trillion-parameter MoE trained entirely on Chinese accelerators to a noncommercial-licensed model raising new questions about US-China AI trade.

DeepSeek has updated its budget model V4 Flash to version '0731,' pushing its Artificial Analysis Intelligence Index score to 50 — just one point behind OpenAI's GPT-5.6 Luna — while costing an estimated 60 percent less per task. The MIT-licensed model keeps its 284B-parameter architecture but shows major gains in agentic benchmarks and token efficiency.

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.

Microsoft has released Mage-VL, a codec-native multimodal foundation model built on a from-scratch 4B-parameter visual encoder paired with Qwen3-4B-Instruct-2507. The model claims up to 3.5x inference speedup over uniform frame sampling and outperforms Qwen3-VL-4B on video and temporal-grounding benchmarks, according to Microsoft.

Liquid AI released two open-weight encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, built for classification and routing tasks at long context on CPU hardware. The company claims the smaller model runs 3.7x faster than ModernBERT-base at 8,192 tokens while matching or beating larger encoders on GLUE, SuperGLUE, and multilingual benchmarks.

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.

Microsoft Research AI Frontiers has released Fara1.5-27B, a 27-billion-parameter multimodal agent that completes web tasks by reading screenshots and emitting click/type/scroll commands. The model, fine-tuned from Qwen3.5-27B, ships under MIT license with a 262K-token context window and is designed to run alongside Microsoft's MagenticLite sandbox.

Microsoft has released Mage-Flow, a 4B-parameter open-weight foundation model for text-to-image generation and instruction-based editing. The company claims it matches or beats much larger open systems like Qwen-Image (20B) and FLUX.2 (32B) while running faster and using less memory.

NVIDIA released Nemotron-Labs-TwoTower-30B-A3B-Base-BF16, a block-wise diffusion language model that generates text by denoising blocks of tokens in parallel rather than sequentially. According to NVIDIA, the model achieves 2.42× the wall-clock generation throughput of its autoregressive baseline while retaining 98.7% of aggregate benchmark quality.

Mistral AI released Medium 3.5, a 128B dense model with a 256k context window that scores 77.6% on SWE-Bench Verified. The model powers new remote coding agents in Mistral Vibe that run asynchronously in the cloud, plus a new Work mode in Le Chat for multi-step agentic tasks.

DeepSeek has released two new models in its V4 series: DeepSeek-V4-Flash and DeepSeek-V4-Pro, both now available on Hugging Face. Separately, Tencent has shipped Hy3-Preview, marking simultaneous releases from two major Chinese AI labs.

Arcee AI has released Trinity Large Preview, a 400B-parameter sparse Mixture-of-Experts model with 13B active parameters per token using 4-of-256 expert routing. The model supports context windows up to 512K tokens and is available with open weights under permissive licensing.