Twelve open-weight language models with 295 billion or more total parameters, ranked by the size each lab states on its model card, with active parameters, context window and license beside it. Kimi K3 is largest at 2.8 trillion, and nine of the twelve activate fewer than 50 billion parameters per token.
Total parameters stated on the model card, ten largest parameters
Vendor-stated As each lab states it on its Hugging Face model card; checked Oct 22, 2026. Bars relative to the largest.
As of All parameter counts vendor-statedLatest checkpoint per model family with a stated count of 295B or more; ungated Hugging Face repositories only
| Rank | Model | Lab | Total parameters (billions) | Relative to #1 | Active per token (billions) | Context window | License, as the card names it | Repository created (Hugging Face) | Figure |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Kimi K321moonshotai/Kimi-K3 | Moonshot AI | 2.8T | 104B | 1M tokens | Kimi K3 License | 2026-06-13 | Vendor-stated | |
| 2 | Qwen3.8-2.4T-A95B31Qwen/Qwen3.8-2.4T-A95B | Alibaba (Qwen) | 2.4T | 95B | 262,144 native; up to 1,010,000 | Qwen3.8-Max license (listed as "other") | 2026-08-08 | Vendor-stated | |
| 3 | DeepSeek-V4-Pro41deepseek-ai/DeepSeek-V4-Pro | DeepSeek | 1.6T | 49B | 1M tokens | MIT | 2026-04-22 | Vendor-stated | |
| 4 | MiMo-V2.6-Pro-RL51XiaomiMiMo/MiMo-V2.6-Pro-RL | Xiaomi | 1.02T | 42B | 1M tokens | MIT | 2026-09-21 | Vendor-stated | |
| 5 | Inkling61thinkingmachines/Inkling | Thinking Machines | 975B | 41B | Not stated on card | Apache 2.0 | 2026-07-14 | Vendor-stated | |
| 6 | Step 5 Preview71stepfun-ai/Step-5-Preview | StepFun | 600B | 27B | Up to 1M tokens | Custom (LICENSE.txt, listed as "other") | 2026-09-20 | Vendor-stated | |
| 7 | Ling-3.1-flash81inclusionAI/Ling-3.1-flash | inclusionAI | 560B | 25B | 1M tokens | MIT | 2026-10-13 | Vendor-stated | |
| 8 | Nemotron-3-Ultra-550B-A55B91nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 | NVIDIA | 550B | 55B | Up to 1M tokens | OpenMDW-1.1 | 2026-06-03 | Vendor-stated | |
| 9 | Beam 1.0 501B-A23B101reflection/Beam-1.0-501B-A23B | Reflection | 501B | 23B | 1M tokens | Apache 2.0 | 2026-10-07 | Vendor-stated | |
| 10 | MiniMax-M3111MiniMaxAI/MiniMax-M3 | MiniMax | ~428B | ~23B | 1M tokens | MiniMax Community (listed as "other") | 2026-06-02 | Vendor-stated | |
| 11 | GLM-5.3-Flash121zai-org/GLM-5.3-Flash | Z.ai | 320B | 18B | Not stated on card | MIT | 2026-08-25 | Vendor-stated | |
| 12 | Hy3131tencent/Hy3 | Tencent | 295B | 21B | 256K tokens | Apache 2.0 | 2026-07-02 | Vendor-stated |
Ranked by total parameters as the lab states them in its model card, from the largest down to 295B; the table shows the latest checkpoint per family and does not list older versions of the same family. Counts are not independently measured and the labs count differently (a few state "about"). "Active per token" is the number of parameters a mixture-of-experts model uses for each token. "Repository created" is the date Hugging Face's API records for the repository, a proxy for the release date and not a vendor statement. License names are quoted as each card gives them; several are custom licenses with their own terms, and this table does not compare them. Left out of the ranked column, for the reasons given below: Meta's Muse Spark Open, Mistral Large 4, GLM-5.3 and DeepSeek-V4.1-Flash. Figure labels: how they work.
An open-weight model is one whose trained weights can be downloaded; the license column shows how varied the terms are, from MIT and Apache 2.0 to custom licenses.1 Here, "open-weight" means a repository on Hugging Face that could be read without an access request when checked on October 22, 2026.1 The ranked figure is the size the lab states, in billions (B) or trillions (T) of parameters. In a mixture-of-experts model most of those parameters sit in experts that a router skips for any given token, so the active count, 18B to 104B in this table, is a better guide to compute per token than the total, which mostly decides how much memory the weights need.14
Moonshot AI's card calls Kimi K3 an "open-weight, native multimodal agentic model" with 2.8T parameters, 104B activated per token, 93 layers and a 1-million-token window; its mixture-of-experts layers activate 16 of 896 experts.2 It is released under the Kimi K3 License.2
Alibaba's Qwen3.8-2.4T-A95B has 2.4T parameters and 95B activated, with 10 routed experts plus one shared, a native context of 262,144 tokens and an extended range to 1,010,000.3 The card says Qwen3.8-Max is the official version built on it, with 1M context by default and added features such as vision input.3
DeepSeek's card describes DeepSeek-V4-Pro as a preview with 1.6T parameters and 49B activated, built for million-token context, under the MIT license.4 A later checkpoint, DeepSeek-V4-Pro-0813, was published on August 13, 2026; its card does not restate the parameter count, so the preview card's figure is used.1
Xiaomi calls MiMo-V2.6-Pro-RL the flagship checkpoint of its MiMo-V2.6 series: a sparse mixture-of-experts model with 1.02T parameters, 42B activated, 384 routed experts of which 8 are active, and a 1M-token window, under the MIT license.5
Thinking Machines' Inkling is a multimodal model that takes text, image and audio input and generates text; its card gives 975B total and 41B active parameters under Apache 2.0 and states no context length.6
StepFun describes Step 5 Preview as a sparse mixture-of-experts model with 600B total and 27B active parameters that handles text, images and video with up to 1 million tokens of context; its license is a custom file.7
Reflection's Beam takes text and generates text; its card gives 501B total and 23B active parameters, a 1M-token context and an Apache 2.0 license, and the release includes BF16, FP8 and NVFP4 versions.10
MiniMax-M3 is, per its card, a native multimodal model with a 1M context, about 428B parameters and about 23B activated, under the MiniMax Community license.11
Meta's Muse Spark Open is described by Hugging Face's launch post as an open-weight model of about 816 billion parameters, around 42 billion active per token, with a context window of up to one million tokens.14 Its Hugging Face repositories were gated with manual approval when checked, so it is not ranked.1 Mistral Large 4 entered public preview on October 6, 2026 as an open-weight model with 1.05T total and 52B active parameters, according to Mistral's documentation; the Hugging Face repository the docs link to returned an access error, so the weights could not be confirmed as downloadable.15 Counted, Mistral Large 4 would rank fourth and Muse Spark Open sixth.
GLM-5.3, from Z.ai, is not ranked because its card states no parameter count; it says it uses the same base model as GLM-5.2, and only its smaller GLM-5.3-Flash is ranked.1612 DeepSeek-V4.1-Flash is not ranked because its card counts differently: 552 billion "backbone" parameters, plus a separate 196 billion-parameter memory component, with 8B and 16B active parameters listed on its card.17 Older members of these families, such as Kimi K2.6, are omitted because a newer one is listed.