Mistral released the model weights of the new version of Le Chonk (Mistral Large 4) on Hugging Face, targeting agentic coding and cybersecurity scenarios. Mistral claims that the model ranks first among open-source models on Cognition's FrontierCode benchmark and sets a new benchmark for Pareto efficiency on Artificial Analysis, while also improving inference efficiency and reducing inference costs. The model is also claimed to achieve SOTA in cybersecurity defense and has been tested by vendors such as CrowdStrike, Cloudflare, Rubrik, and Zscaler; users can customize and self-deploy it with Mistral Forge.
Mistral released the model weights of the new version of Le Chonk (Mistral Large 4) on Hugging Face, targeting agentic coding and cybersecurity scenarios. Mistral claims that the model ranks first among open-source models on Cognition's FrontierCode benchmark and sets a new benchmark for Pareto efficiency on Artificial Analysis, while also improving inference efficiency and reducing inference costs. The model is also claimed to achieve SOTA in cybersecurity defense and has been tested by vendors
After a three-week preview, Mistral AI has released the weights of Mistral Large 4 on Hugging Face. This multimodal model was trained on 3,800 Nvidia accelerators and has a parameter count of one trillion, and it is currently still difficult to run on a gaming PC or a single GPU. Mistral says it leads GLM-5.3 and Kimi K3 on the cybersecurity benchmark CyberGym-E2E, and is on par with GLM-5.3 on the agentic coding benchmark Terminal Bench 4.0, but there is no independent evaluation yet; preview pricing is $1.4 per million input tokens and $4.2 per million output tokens.
After a three-week preview, Mistral AI has released the weights of Mistral Large 4 on Hugging Face. This multimodal model was trained on 3,800 Nvidia accelerators and has a parameter count of one trillion, and it is currently still difficult to run on a gaming PC or a single GPU. Mistral says it leads GLM-5.3 and Kimi K3 on the cybersecurity benchmark CyberGym-E2E, and is on par with GLM-5.3 on the agentic coding benchmark Terminal Bench 4.0, but there is no independent evaluation yet; preview p
Mistral AI has released the model weights for the new version of Le Chonk, claiming it ranks first among open models on Cognition's FrontierCode and sets a new benchmark on Artificial Analysis's Pareto efficiency. The model is described as reaching SOTA in cybersecurity defense and has been tested by vendors including CrowdStrike, Cloudflare, Rubrik, and Zscaler. The weights are now available for download on Hugging Face and can be customized and self-deployed with Mistral Forge.
Mistral AI announced the open release of the model weights for a new version of Le Chonk, saying it advances the open frontier in agentic coding and cybersecurity. The model ranks first among open models on Cognition's FrontierCode and sets a new benchmark for Pareto efficiency on Artificial Analysis; on cybersecurity it claims state-of-the-art, having been tested by vendors including CrowdStrike, Cloudflare, Rubrik, and Zscaler. The weights can be downloaded on Hugging Face and customized and deployed with Mistral Forge. Arthur Mensch reposted and said progress over recent weeks has been good, and now the weights are open.
gentic coding and cybersecurity. The model ranks first among open models on Cognition's FrontierCode and sets a new benchmark for Pareto efficiency on Artificial Analysis; on cybersecurity it claims state-of-the-art, having been tested by vendors including CrowdStrike, Cloudflare, Rubrik, and Zscaler. The weights can be downloaded on Hugging Face and customized and deployed with Mistral Forge. Arthur Mensch reposted and said progress over recent weeks has been good, and now the weights are open.
Mistral has released the model weights of its flagship model Large 4 (ML4), which was previously available only as an API preview. After three weeks of testing by customers and cybersecurity partners and further training with reinforcement learning, ML4's success rate on Terminal-Bench 4.0 rose from 28% to 41%, and two new reasoning effort levels, Low and Medium, were added. Mistral research lead Pierre Stock said that at equivalent capability the new version uses about 50% fewer tokens than the preview; the weights became available for download on Hugging Face at 19:00 and are also offered through the Mistral API and the coding assistant Vibe.
testing by customers and cybersecurity partners and further training with reinforcement learning, ML4's success rate on Terminal-Bench 4.0 rose from 28% to 41%, and two new reasoning effort levels, Low and Medium, were added. Mistral research lead Pierre Stock said that at equivalent capability the new version uses about 50% fewer tokens than the preview; the weights became available for download on Hugging Face at 19:00 and are also offered through the Mistral API and the coding assistant Vibe.
On October 28, 2026, NVIDIA released the speech transcription and translation model Nemotron 3.5 Transcribe, supporting 40 languages including Japanese and capable of transcribing recordings up to 1 hour long. The model is an open-weight 30B MoE architecture that activates approximately 4B parameters during inference, can specify via text instructions whether to retain filler words and how numbers are written, and supports speaker diarization for up to 8 people and bidirectional Japanese-English translation. The weights are released on Hugging Face as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B under the OpenMDW-1.1 license, with BF16, FP8, and NVFP4 versions and NeMo Speech fine-tuning tools provided.
ng. The model is an open-weight 30B MoE architecture that activates approximately 4B parameters during inference, can specify via text instructions whether to retain filler words and how numbers are written, and supports speaker diarization for up to 8 people and bidirectional Japanese-English translation. The weights are released on Hugging Face as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B under the OpenMDW-1.1 license, with BF16, FP8, and NVFP4 versions and NeMo Speech fine-tuning tools provided.
Mistral released the weights of its frontier open-source model Mistral Large 4 (ML4) on Hugging Face on Thursday; the model had only been available in preview three weeks earlier. Mistral calls ML4 "the world's best open-source agentic coding model," scoring 41% on Terminal-Bench 4.0, slightly above GLM-5.3's 40%, and says it uses an average of 47% fewer tokens per task than GLM-5.3; it scored 79% on the CyberGym E2E defensive cybersecurity test, compared with 29% for GLM-5.3. Mistral Vice President of Science Pierre Stock called ML4 merely a "stepping stone" and confirmed that smaller versions will be released in the coming months, one of which can run on a laptop.
Mistral released the weights of its frontier open-source model Mistral Large 4 (ML4) on Hugging Face on Thursday; the model had only been available in preview three weeks earlier. Mistral calls ML4 "the world's best open-source agentic coding model," scoring 41% on Terminal-Bench 4.0, slightly above GLM-5.3's 40%, and says it uses an average of 47% fewer tokens per task than GLM-5.3; it scored 79% on the CyberGym E2E defensive cybersecurity test, compared with 29% for GLM-5.3. Mistral Vice Presi
Mistral AI released the open weights of Mistral Large 4 (codename Le Chonk) on Hugging Face on October 29. The model has 100 billion parameters out of a total parameter scale of the 100-billion class, with about 4.9 billion parameters actually invoked per token, using a MoE architecture with 128 experts, of which 4 are activated each time plus 1 shared expert. Compared with the October 6 preview version, Mistral says it continued reinforcement learning over three weeks, raising the Terminal-Bench 4.0 score from 28% to 41%; the Artificial Analysis coding index is 48%, on par with GLM-5.3 while using an average of 47% fewer tokens per task. API pricing drops to $0.70 per million input tokens and $2.10 per million output tokens; the new Mistral Ecosystem Open-Weight License permits commercial use, but reselling model services requires a separate agreement, and the full version requires 8 Nvidia B200 or B300 GPUs.
Mistral AI released the open weights of Mistral Large 4 (codename Le Chonk) on Hugging Face on October 29. The model has 100 billion parameters out of a total parameter scale of the 100-billion class, with about 4.9 billion parameters actually invoked per token, using a MoE architecture with 128 experts, of which 4 are activated each time plus 1 shared expert. Compared with the October 6 preview version, Mistral says it continued reinforcement learning over three weeks, raising the Terminal-Benc
MiniMax has publicly released the MiniMax M3.1 Flash weights on Hugging Face; the model has 428B total parameters and 23B activated per token. The company claims generational improvements in agent tasks, coding, and cybersecurity, with scores of 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82.2 on CyberGym. This model is the stealth model Space Bunny on OpenRouter, released on October 19, and is now available on the MiniMax platform and API.
Thomas Wolf announced that, in collaboration with TIGEM, the University of Naples Federico II, and Zhongguancun Academy, they trained Carbon-X, an open 14B model that can directly read raw DNA and interpret it in natural language. It scores 67% on its DNA benchmark, above Claude Opus 5.5's 61% and Gemini 4 Argon's 50%, and is less than one point behind GPT-6.1 Sol while costing about 1,000 times less. The training data excludes human and animal viruses, the weights are open, and the model can run on hardware accessible to academia, such as a laptop.
Thomas Wolf announced that, in collaboration with TIGEM, the University of Naples Federico II, and Zhongguancun Academy, they trained Carbon-X, an open 14B model that can directly read raw DNA and interpret it in natural language. It scores 67% on its DNA benchmark, above Claude Opus 5.5's 61% and Gemini 4 Argon's 50%, and is less than one point behind GPT-6.1 Sol while costing about 1,000 times less. The training data excludes human and animal viruses, the weights are open, and the model can ru
Hugging Face co-founder Clément Delangue announced the open-source DNA model Carbon-X, obtained by mid-training Qwen3-14B on 300B tokens of DNA, English, and interleaved data, using a hybrid tokenizer with BPE for English and a 6-mer tokenizer for DNA. The model surpasses all open-source language models it evaluated on five DNA understanding and generation benchmarks, improving on Qwen3-14B by 27 points, with API costs 1,800 times lower than GPT 6.1 Sol. The model is released on Hugging Face under the Apache 2.0 license.
Hugging Face co-founder Clément Delangue announced the open-source DNA model Carbon-X, obtained by mid-training Qwen3-14B on 300B tokens of DNA, English, and interleaved data, using a hybrid tokenizer with BPE for English and a 6-mer tokenizer for DNA. The model surpasses all open-source language models it evaluated on five DNA understanding and generation benchmarks, improving on Qwen3-14B by 27 points, with API costs 1,800 times lower than GPT 6.1 Sol. The model is released on Hugging Face und
Hugging Face on Thursday released the open-source DNA understanding model Carbon-X, which can run locally on a laptop, letting users ask in natural language how DNA changes affect gene activity and health. The model was developed in collaboration with researchers at Zhongguancun Academy and the University of Naples Federico II; it extends the vocabulary of an existing language model to read DNA in six-letter units and was further trained on 300 billion tokens of DNA sequences, English text, and samples pairing DNA with biological descriptions. The model has 14B parameters and can process sequences of about 50,000 base pairs at a time; it cannot upload a complete human genome. The developers also note that the prediction accuracy has not yet been validated through wet-lab experiments.
Hugging Face on Thursday released the open-source DNA understanding model Carbon-X, which can run locally on a laptop, letting users ask in natural language how DNA changes affect gene activity and health. The model was developed in collaboration with researchers at Zhongguancun Academy and the University of Naples Federico II; it extends the vocabulary of an existing language model to read DNA in six-letter units and was further trained on 300 billion tokens of DNA sequences, English text, and
Hugging Face has released the open-source model Carbon-X, combining DNA analysis and generation with natural language so that researchers can ask questions directly and receive predictions and explanations. The model is based on Qwen3-14B and was trained on genomic sequences and natural language from humans, other mammals, plants, and fungi, with viruses explicitly excluded from the training data. Researchers can run it locally, avoiding the sharing of sensitive patient data with external AI companies, and can also generate new DNA sequences to assist gene therapy research.
Hugging Face has released the open-source model Carbon-X, combining DNA analysis and generation with natural language so that researchers can ask questions directly and receive predictions and explanations. The model is based on Qwen3-14B and was trained on genomic sequences and natural language from humans, other mammals, plants, and fungi, with viruses explicitly excluded from the training data. Researchers can run it locally, avoiding the sharing of sensitive patient data with external AI com
Hugging Face has released Beam, a 501B-parameter sparse MoE open source model aimed at long-text reasoning and agent tasks with 1M context, activating only about 23B parameters per token. The model uses 52 layers of hybrid attention, with 43 layers using a 512-token sliding window and 9 layers using global attention; the KV cache grows by 36 KiB per token, about one-sixth that of a full-attention approach. The official release provides three checkpoints—BF16, FP8, and NVFP4; NVFP4 can compress the weights from about 1 TB to about 280 GB, and day-0 support is already available in transformers, vLLM, and SGLang.
Hugging Face has released Beam, a 501B-parameter sparse MoE open source model aimed at long-text reasoning and agent tasks with 1M context, activating only about 23B parameters per token. The model uses 52 layers of hybrid attention, with 43 layers using a 512-token sliding window and 9 layers using global attention; the KV cache grows by 36 KiB per token, about one-sixth that of a full-attention approach. The official release provides three checkpoints—BF16, FP8, and NVFP4; NVFP4 can compress t
Meta has released the model weights of Muse Spark Open on Hugging Face, an 816B-parameter model aimed at coding and agent tasks. The open-source version shares the same base model as Muse Spark 1.3, with additional safety post-training to reduce the risk of misuse when used standalone without system guardrails, reversing Meta's strategy of its first proprietary closed frontier model in April this year. The model has 62 layers, each containing 256 expert networks, with each token routed to 8 of them, so deployment requires the memory of an 816B-parameter model but runs at the speed of a 42B model. A 3-bit quantized version tested on a 512GB M3 Ultra occupies about 344 GB, with generation speed of about 21 tok/s under short prompts.
Meta has released the model weights of Muse Spark Open on Hugging Face, an 816B-parameter model aimed at coding and agent tasks. The open-source version shares the same base model as Muse Spark 1.3, with additional safety post-training to reduce the risk of misuse when used standalone without system guardrails, reversing Meta's strategy of its first proprietary closed frontier model in April this year. The model has 62 layers, each containing 256 expert networks, with each token routed to 8 of t
On October 29, Baseten released joey-1, a 2B-parameter decision model, priced at $0.01 per million tokens. It is now available on Baseten Model APIs and has open weights on Hugging Face. The model was post-trained by the in-house lab Base Labs based on Qwen3.5-2B, and shares the same data recipe and API with mammoth-1 (27B) and elephant-1 (9B), so switching requires only a one-line change. The model card shows that joey-1 scored 63.30 on the public Jev Decision Index 0.4 leaderboard, ranking 37th among 230 models and the highest score among models with fewer than 4B parameters; Baseten says its P50 end-to-end latency for 2000-token requests in production is 20 ms.
On October 29, Baseten released joey-1, a 2B-parameter decision model, priced at $0.01 per million tokens. It is now available on Baseten Model APIs and has open weights on Hugging Face. The model was post-trained by the in-house lab Base Labs based on Qwen3.5-2B, and shares the same data recipe and API with mammoth-1 (27B) and elephant-1 (9B), so switching requires only a one-line change. The model card shows that joey-1 scored 63.30 on the public Jev Decision Index 0.4 leaderboard, ranking 37t
Nvidia has released the open-weight multilingual transcription model Nemotron 3.5 Transcribe, which ranks first on the Hugging Face Open ASR English open-source model leaderboard with an average word error rate of 5.33%. The model uses a 30B MoE architecture with 4B activated parameters, supports transcription in 40 languages and translation between 31 languages and English, can label speakers and distinguish up to 8 voices, handles recordings up to 1 hour long, but does not support real-time streaming. Users can provide instructions such as context and terminology before transcription; the model can be deployed with vLLM or run locally on a single GPU via NeMo-Speech.cpp; Nvidia has also released the data curation pipeline Granary v2 and a public manifest of approximately 463,062 hours across 48 languages.
Nvidia has released the open-weight multilingual transcription model Nemotron 3.5 Transcribe, which ranks first on the Hugging Face Open ASR English open-source model leaderboard with an average word error rate of 5.33%. The model uses a 30B MoE architecture with 4B activated parameters, supports transcription in 40 languages and translation between 31 languages and English, can label speakers and distinguish up to 8 voices, handles recordings up to 1 hour long, but does not support real-time st