Mistral Opens Mistral Large 4 Preview, Promises Open Weights in October
Mistral opened a public preview of Mistral Large 4 on October 6, 2026. The French company says the 1 trillion-parameter model's weights will be released by the end of October.
// ai models
Open-Weight AI Model Releases Tracker
A dated, sourced timeline of large open-weight AI model releases from any lab since January 2025. Each entry marks the day the weights became downloadable, with license and size where stated.
Researched and fact-checked by AI, with no human review. How we verify
26 entries from 32 sources. Checked for news weekly; last checked
What this tracker follows
This tracker follows the day an AI lab makes the weights of a large general-purpose model downloadable. Weights are the numbers a model learns in training. Publishing them lets anyone run the model on their own hardware. The tracker covers any lab from January 2025 on. It is current as of October 17, 2026.
An entry marks the first day the weights of a lab's new flagship open model could be downloaded. The model must work with text, alone or with images and other media. It must have at least 100 billion parameters. The lab must also present it as its most capable open release. A target date that a lab has named gets a scheduled entry.
Entries give the license and the parameter counts where a cited source states them. Most of these models are mixture-of-experts designs. They use only part of the network for each token, a small piece of text. That is why a total count and an active count both appear. Performance claims are the labs' own.
Smaller or specialized versions, date-stamped updates, fine-tunes and re-uploads are left out. So are image, video, audio and embedding models. The list is a selection that favors widely covered releases. Dates follow the cited source, so a release in China can carry a date one day later than in a U.S. report.
The site's AI model release tracker is different. It follows major model launches from eight labs, open or closed. This tracker follows only downloadable weights, from any lab, with license and size.
Scheduled next
Targets set by the organizations named. Dates like these often move.
Scheduled
Mistral targets the end of October for Mistral Large 4 open weights
Mistral said on October 6, 2026 that it would release the weights by the end of the month. On October 17, its Hugging Face release page gave October 31 as the current estimate. The page describes a 1 trillion-parameter multimodal model with 49 billion active parameters, or 52 billion counting embeddings and output layers.
Source: mistralai/Mistral-Large-4-1T-A52B · Upcoming release, Mistral AI, via Hugging Face
Timeline
Reflection AI releases Beam, its first open-weight model
Reflection AI released Beam's weights on Hugging Face under the Apache 2.0 license. Beam is a mixture-of-experts model with 501 billion total parameters, 23 billion of them active. Reflection says it is built for coding, reasoning and long agent tasks.
Meta releases Muse Spark Open, an open-weight version of Muse Spark
Meta released Muse Spark Open, an 816 billion-parameter mixture-of-experts model that activates about 42 billion parameters per token. It takes text, image, video and audio input. Hugging Face hosts the checkpoints, and Meta's own Muse Model License applies.
Source: Build with Muse Spark Open, Meta
Xiaomi open-sources MiMo-V2.6 Pro and MiMo-V2.6 Flash
Xiaomi released and open-sourced its MiMo-V2.6 series, with weights for the Pro and Flash models on Hugging Face. It also released a smaller distilled model. Xiaomi says Pro scores 46 on the Artificial Analysis Intelligence Index, making it the most powerful open-source model available.
Source: MiMo-V2.6: Scaling Up Reinforcement Learning for Self-Improvement, Xiaomi
DeepSeek releases DeepSeek-V4.1-Flash with open weights
DeepSeek released V4.1-Flash, a 552 billion-parameter mixture-of-experts model. It uses 8 billion active parameters for input and 16 billion for output. The release post links the weights on Hugging Face. DeepSeek says its benchmark results beat V4-Pro.
Source: DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient, DeepSeek
Z.ai releases GLM-5.3 weights under a new custom license
Z.ai put the weights of GLM-5.3, a 753 billion-parameter mixture-of-experts model, on Hugging Face. Z.ai says it held them back two weeks for safety testing. The new GLM-5.3 License replaces MIT. Companies that host it and earn over $10 billion in any 12 months must pass a Z.ai security review for commercial use.
Source: Z.ai’s GLM-5.3 goes open weight, but its new license aims at hyperscalers, The New Stack
Alibaba's Qwen team releases Qwen3.8-2.4T-A95B open weights
The Qwen team said Qwen3.8-2.4T-A95B was available on Hugging Face and ModelScope. The model card lists 2.4 trillion total and 95 billion activated parameters, under a license named qwen3.8-max. A smaller model, Qwen3.8-27B, followed on August 14.
Moonshot AI releases the full Kimi K3 weights
Moonshot released the weights of Kimi K3, which had launched on its hosted API earlier in July. VentureBeat describes a 2.8 trillion-parameter model that activates 104 billion. It ships under a custom Kimi K3 License with conditions for large commercial users.
Thinking Machines Lab releases Inkling with open weights
The model card dates Inkling's release to July 15, 2026, under the Apache 2.0 license, with weights on Hugging Face. It is a mixture-of-experts model with 975 billion total and 41 billion active parameters. It accepts text, image and audio input.
Z.ai releases GLM-5.2 with open weights under the MIT license
Z.ai released GLM-5.2, a 753 billion-parameter model, with weights available at once on Hugging Face under the MIT license. VentureBeat reported a 1 million-token context window and a focus on long-running coding tasks.
Nvidia releases Nemotron 3 Ultra as an open model
Nvidia's model card gives a Hugging Face release date of June 4, 2026 for Nemotron 3 Ultra. It has 550 billion total and 55 billion active parameters in a hybrid Mamba-Transformer mixture-of-experts design. It uses the OpenMDW 1.1 license.
Source: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16, NVIDIA, via Hugging Face
DeepSeek releases DeepSeek-V4 preview with open weights
DeepSeek released a preview of V4 with open weights on Hugging Face and same-day API access. V4-Pro has 1.6 trillion total and 49 billion active parameters. V4-Flash has 284 billion total and 13 billion active. Both support a 1 million-token context.
Moonshot AI open-sources Kimi K2.6
Moonshot said it was open-sourcing Kimi K2.6, its latest model, with a focus on coding, long-horizon tasks and agent swarms. Moonshot's research index dates the post April 20, 2026. The K2.6 model card on Hugging Face lists 1 trillion total and 32 billion activated parameters.
Source: Kimi K2.6: Advancing Open-Source Coding, Moonshot AI
Z.ai releases GLM-5.1 under the MIT license
Z.ai released GLM-5.1 with weights on Hugging Face and ModelScope under the MIT license. VentureBeat describes a 754 billion-parameter mixture-of-experts model with a 202,752-token context window. Z.ai says it can work on one task for up to eight hours.
Alibaba's Qwen team releases Qwen3.5, led by a 397B model
The Qwen team released Qwen3.5, starting with a mixture-of-experts model, Qwen3.5-397B-A17B. The model card lists 397 billion total and 17 billion activated parameters. More sizes followed on February 24 and March 2, down to Qwen3.5-0.8B.
MiniMax releases MiniMax M2.5 as open weights
MiniMax released M2.5 on Hugging Face under a modified MIT license. Commercial users must display the MiniMax M2.5 name in their product. VentureBeat reports 230 billion total parameters, with 10 billion active.
Z.ai releases GLM-5 with open weights under the MIT license
VentureBeat reported that Z.ai's new GLM-5 keeps the MIT license and is available as open weights. It has 744 billion total parameters, 40 billion active per token, up from 355 billion in GLM-4.5. It supports a 200,000-token context. Z.ai's own release notes date the launch February 12.
Moonshot AI releases Kimi K2.5 as an open-source model
Moonshot released Kimi K2.5, which takes text and image input, under a Modified MIT license. Products with over 100 million monthly users or $20 million in monthly revenue must display its name. The K2.5 model card on Hugging Face lists 1 trillion total and 32 billion activated parameters.
Source: How Moonshot's Kimi K2.5 helps AI builders spin up agent swarms easier than ever, VentureBeat
Mistral releases Mistral Large 3 under the Apache 2.0 license
Mistral released Mistral 3, led by Mistral Large 3, a mixture-of-experts model with 675 billion total and 41 billion active parameters. Three small Ministral 3 models came with it. All are under the Apache 2.0 license, with weights on Hugging Face.
DeepSeek releases DeepSeek-V3.2 and DeepSeek-V3.2-Speciale
DeepSeek released V3.2, the successor to V3.2-Exp, and the reasoning-focused V3.2-Speciale. The release post links open weights for both on Hugging Face. On DeepSeek's own service, Speciale was API-only at first.
Source: DeepSeek-V3.2 Release, DeepSeek
MiniMax open-sources MiniMax M2
MiniMax released M2, a model built for coding and agent tasks, and put the full weights on Hugging Face. MiniMax said its API price was 8% of Claude Sonnet's, at about twice the speed.
Source: MiniMax M2 & Agent: Ingenious in Simplicity, MiniMax
OpenAI releases gpt-oss-120b and gpt-oss-20b
OpenAI released two open-weight models under the Apache 2.0 license, its first open-weight language models since GPT-2. The larger, gpt-oss-120b, has 117 billion total parameters with 5.1 billion active. The weights are on Hugging Face.
Source: Introducing gpt-oss, OpenAI
Moonshot AI open-sources Kimi K2
Moonshot open-sourced Kimi K2 in base and instruction-tuned versions. It is a mixture-of-experts model with 1 trillion total parameters and 32 billion active. Moonshot says it is optimized for agent tasks such as tool use. Moonshot's research index dates the post July 11, 2025.
Alibaba's Qwen team releases Qwen3 as open weights
The Qwen team released open weights for two mixture-of-experts models and six dense models. The largest, Qwen3-235B-A22B, has 235 billion total and 22 billion active parameters. The dense models are under the Apache 2.0 license.
Source: Qwen3: Think Deeper, Act Faster, Qwen team, Alibaba Group
Meta releases Llama 4 Scout and Llama 4 Maverick
Meta released Llama 4 Scout and Llama 4 Maverick for download on llama.com and Hugging Face. Maverick has 400 billion total parameters and Scout has 109 billion. Each has 17 billion active parameters.
Source: The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation, Meta
DeepSeek releases DeepSeek-R1 under the MIT license
DeepSeek released R1, a reasoning model it says performs on par with OpenAI's o1, as a fully open-source model under the MIT license. It also open-sourced six smaller models distilled from R1.
Source: DeepSeek-R1 Release, DeepSeek
Additional sources
- Introducing Mistral Large 4, Mistral AI
- New Released, Z.ai
- Kimi Research & Tech Blog, Moonshot AI
- moonshotai/Kimi-K2.6, Moonshot AI, via Hugging Face
- moonshotai/Kimi-K2.5, Moonshot AI, via Hugging Face
- Qwen/Qwen3.8-2.4T-A95B, Qwen team, Alibaba Group, via Hugging Face
- Qwen/Qwen3.5-397B-A17B, Qwen team, Alibaba Group, via Hugging Face
Update history
- A new tracker logs the day AI labs made large open-weight models downloadable, starting in January 2025. Entries give the license and size where a cited source states them.
Articles on AI Models
ChatGPT, Gemini, Claude and Grok: How the Leading AI Models Differ
ChatGPT, Gemini, Claude and Grok differ on price, access and specialty, and no one model leads every test. Current versions, costs and benchmark claims as of October 2026.
Open-Weight vs Closed AI Models: What the Difference Means for You
Open-weight models can be downloaded and run on your own hardware, while closed models stay on their maker's servers. Open-weight is not the same as open source, and licenses vary.