Tech & AI

The Largest Open-Weight AI Language Models

The biggest AI language models whose weights anyone can download, one per lab, ranked by total parameters as of October 17, 2026. 2026 brought a run of trillion-parameter releases, led by Moonshot AI's 2.8-trillion-parameter Kimi K3.

Video thumbnail for Kimi K3, Moonshot AI's open-weight language model

Pictured: video thumbnail for Kimi K3 from Moonshot AI's Kimi channel, #1 with 2.8 trillion parameters.

Rank Model Total parameters Active per token Lab Country Weights released License Architecture
1Kimi K32.8T104BMoonshot AIChinaJul 2026Kimi K3 LicenseMoE
2Qwen3.8-2.4T-A95B2.4T95BAlibaba (Qwen)ChinaAug 2026Qwen custom licenseMoE
3DeepSeek-V4-Pro1.6T49BDeepSeekChinaApr 2026MITMoE
3LongCat-2.01.6T48BMeituanChinaJul 2026MITMoE
5MiMo-V2.5-Pro1.02T42BXiaomiChinaApr 2026MITMoE
6Yuan3.0 Ultra1.01T68.8BYuanLab (Inspur)ChinaMar 2026Yuan model licenseMoE, multimodal
7Ling-2.5-1T1T63BAnt GroupChinaFeb 2026MITMoE
8Muse Spark Open816B42BMetaUnited StatesOct 2026Muse Model LicenseMoE
9Hy4 preview770B49BTencentChinaAug 2026Apache 2.0MoE
10K-EXAONE 2.0750B37BLG AI ResearchSouth KoreaJul 2026Apache 2.0MoE
11GLM-5744B40BZ.ai (Zhipu)ChinaFeb 2026MITMoE
12openPangu-Ultra-MoE-718B718B39BHuaweiChinaOct 2025openPangu licenseMoE
13A.X K2688B33BSK TelecomSouth KoreaJul 2026Apache 2.0MoE
14Mistral Large 3675B41BMistral AIFranceDec 2025Apache 2.0MoE
15Step 5 Preview600B27BStepFunChinaOct 2026StepFun Model LicenseMoE
16Nemotron 3 Ultra550B55BNvidiaUnited StatesJun 2026OpenMDW-1.1Hybrid Mamba-MoE
17Snowflake Arctic480B17BSnowflakeUnited StatesApr 2024Apache 2.0Dense-MoE hybrid
18MiniMax-M1456B45.9BMiniMaxChinaJun 2025Apache 2.0Hybrid MoE
19ERNIE-4.5-VL-424B-A47B424B47BBaiduChinaJun 2025Apache 2.0MoE, multimodal
20Trinity Large398B13BArcee AIUnited StatesJan 2026Apache 2.0MoE
20Jamba 1.5 Large398B94BAI21 LabsIsraelAug 2024Jamba Open Model LicenseHybrid Mamba-MoE

Basis and caveats: each lab's largest open-weight language model (downloadable weights on Hugging Face, ModelScope or GitCode, under any license), ranked by the total parameter count the lab states, as of October 17, 2026. One row per lab, so a lab's older or smaller releases aren't listed separately. Only base models a lab trained itself count: fine-tunes of another lab's model, models that are API-only (such as Baidu's ERNIE 5) and weights that are promised but not yet out (xAI's Grok 3, Reflection AI's Beam, Meta's Llama 4 Behemoth) are left out. Some checkpoints hold slightly more parameters than the stated figure because they bundle extra modules used to speed up generation; the stated figure is used. Almost every model here is a mixture of experts, which only switches on part of the network for each word, so the “Active per token” column is a better guide to the computing power each one needs to run. “Permissive” means MIT, Apache 2.0 or OpenMDW; “Custom” licenses add conditions such as revenue caps or use restrictions. The table stops at 398 billion parameters; just below it are xAI's Grok-1 (314 billion), TII's Falcon 180B, BigScience's BLOOM (176 billion) and OpenAI's gpt-oss-120b (117 billion). New releases arrive almost monthly, so the order will change.[16]

The Top 10

#1

Kimi K3 (Moonshot AI) — 2.8 trillion parameters

Moonshot AI put Kimi K3 live in its apps and API on July 16, 2026, then published the full weights on Hugging Face at the end of the month: 96 shards totalling about 1.56 TB. Its model card lists 2.8 trillion parameters spread over 896 experts, 16 of which fire for each token, for 104 billion active parameters. The weights ship in a 4-bit format the model was trained to tolerate. The Kimi K3 License allows commercial use and modification, but a company selling the model as a service with more than $20 million in revenue over 12 months needs a separate agreement. Moonshot's previous record holder, Kimi K2 (July 2025), had about 1 trillion parameters.[1][2]

#2

Qwen3.8-2.4T-A95B (Alibaba) — 2.4 trillion

Alibaba first offered Qwen3.8-Max through its API in early August 2026, then published the weights on August 12 as Qwen3.8-2.4T-A95B: 2.4 trillion parameters with 95 billion active. It is the first Max-class Qwen anyone can download. The open checkpoint is a text-only reasoning model, separate from the multimodal model behind the API, and comes under a custom license rather than the Apache 2.0 used for smaller Qwens. Before it, the biggest open Qwen was Qwen3.5-397B-A17B from February 2026.[3][4]

#3

DeepSeek-V4-Pro (DeepSeek) — 1.6 trillion

DeepSeek released V4 as a preview on April 24, 2026, in two sizes: V4-Pro at 1.6 trillion parameters with 49 billion active, and V4-Flash at 284 billion with 13 billion active. Both have a 1-million-token context window and use the MIT license, so anyone can use them commercially. V4-Pro is more than twice the size of the 685-billion-parameter V3.2 it replaced, and DeepSeek shipped an updated V4-Pro checkpoint in August.[5][6]

#3

LongCat-2.0 (Meituan) — 1.6 trillion

LongCat-2.0 comes from Meituan, China's biggest food-delivery company. It matches DeepSeek-V4-Pro at 1.6 trillion parameters, with about 48 billion active per token and a 1-million-token context window. Before Meituan revealed it at the end of June 2026, it had spent two months on OpenRouter under the anonymous name "Owl Alpha". Meituan says it trained the model on more than 50,000 Chinese-made AI chips rather than Nvidia GPUs. The weights followed in July under the MIT license.[7][8]

#5

MiMo-V2.5-Pro (Xiaomi) — 1.02 trillion

Yes, the phone maker. Xiaomi's first trillion-parameter model, MiMo-V2-Pro, launched in March 2026 as an API-only product. A month later the company open-sourced MiMo-V2.5 and MiMo-V2.5-Pro under the MIT license. The Pro's Hugging Face index counts 1,023,244,718,976 parameters, with 42 billion active per token and a 1-million-token context window.[9]

#6

Yuan3.0 Ultra (YuanLab) — 1.01 trillion

Yuan3.0 Ultra comes from YuanLab, the AI arm of Chinese server maker Inspur, and was released on March 4, 2026. It can read images as well as text, and it is 1.01 trillion parameters with 68.8 billion active. It started out bigger: the team began with a 1.515-trillion-parameter design and cut it down to 1.01 trillion during pre-training by pruning experts that weren't pulling their weight, which they say made training 49% more efficient.[10]

#7

Ling-2.5-1T (Ant Group) — 1 trillion

Ant Group, the company behind Alipay, released Ling-2.5-1T in February 2026 alongside a reasoning-focused sibling, Ring-2.5-1T. Ling-2.5-1T has 1 trillion parameters, 63 billion of them active per token, and handles up to 1 million tokens of context. Ant reported gold-medal-level scores for Ring-2.5-1T on the 2025 International Mathematical Olympiad (35 of 42). Ant had already open-sourced Ring-1T in October 2025, billed at the time as the first open trillion-parameter reasoning model.[11]

#8

Muse Spark Open (Meta) — 816 billion

Meta's newest entry landed two days before this ranking: on October 15, 2026, Meta published the weights of Muse Spark Open, an 816-billion-parameter model with about 42 billion active per token and up to 1 million tokens of context. It is built on the August release of Muse Spark 1.2, with extra safety training. It isn't open source in the usual sense: the custom Muse Model License governs its use. Meta tuned it to run locally on two Apple Mac Studios through Apple's MLX framework. It is Meta's biggest open model, ahead of Llama 4 Maverick (400 billion) and the dense Llama 3.1 405B.[12]

#9

Hy4 preview (Tencent) — 770 billion

Tencent published the Hy4 preview on August 28, 2026: 770 billion parameters, 49 billion active, 256 experts plus one shared expert across 78 layers, and a 1-million-token context window. The full weights take about 1.56 TB. Unlike some earlier Hunyuan models, which came with Tencent's own usage terms, it ships under the standard Apache 2.0 license. Tencent's previous record holder, Hunyuan-Large (389 billion) from 2024, was billed at its release as the largest open mixture-of-experts model.[13]

#10

K-EXAONE 2.0 (LG AI Research) — 750 billion

LG's AI lab released K-EXAONE 2.0 on Hugging Face on July 31, 2026, under the Apache 2.0 license. It has 750 billion parameters with 37 billion active, 256 experts plus one shared expert, and a 262,144-token context window. It supports ten languages, up from six in the 236-billion-parameter first K-EXAONE. It was built for South Korea's government-backed sovereign AI program, which also produced #13, SK Telecom's 688-billion-parameter A.X K2.[14][15]

AI Open Source Language Models Technology

Sources

  1. Moonshot AI, Kimi K3 model card and license — huggingface.co/moonshotai/Kimi-K3; The Standard, "Moonshot AI releases full model weights of its Kimi K3, world's largest open-weight AI model" — thestandard.com.hk
  2. Pasquale Pillitteri, "Kimi K3 Open Weights: 104B Active of 2.8T" (technical report summary) — pasqualepillitteri.it
  3. Qwen, "Qwen3.8-Max: A New Bar for Coding and Cowork" — qwen.ai; Qwen3.8-2.4T-A95B weights — huggingface.co/Qwen/Qwen3.8-2.4T-A95B
  4. LLM Stats, "Qwen3.8-Max Open Weights: First Max-Class Qwen You Can Download" — llm-stats.com; Artificial Analysis, Qwen3.5-397B-A17B open weights release — artificialanalysis.ai
  5. DeepSeek, DeepSeek-V4-Pro model card — huggingface.co/deepseek-ai
  6. Let's Data Science, "DeepSeek V4 Preview Adds Open-Weight 1M-Context Models" — letsdatascience.com; NYU Shanghai RITS, "DeepSeek Releases V4: Open-Source 1.6T MoE with 1M Context" — rits.shanghai.nyu.edu
  7. Meituan, LongCat-2.0 model card — huggingface.co/meituan-longcat
  8. ComfyUI Wiki, "Meituan Open Sources LongCat 2.0: 1.6 Trillion Parameter MoE Model Trained on AI ASICs" — comfyui-wiki.com; Digital Today, "Meituan releases LongCat 2.0 open source, trained on 50,000 Chinese ASICs" — digitaltoday.co.kr
  9. Korben, "Xiaomi open-sources MiMo-V2.5 and MiMo-V2.5-Pro under MIT license" — korben.info; Wikipedia, "Xiaomi MiMo" — en.wikipedia.org
  10. YuanLab, "Yuan3.0 Ultra: A Trillion-Parameter Enterprise-Oriented MoE LLM" (arXiv 2601.14327) — arxiv.org; Yuan3.0 Ultra model card — huggingface.co/YuanLabAI
  11. Ant Group, "Ant Group Releases Ling 2.5 1T and Ring 2.5 1T" (Business Wire, Feb 2026) — businesswire.com; European Open Source AI Index, Ling — osai-index.eu
  12. SiliconANGLE and The New Stack on Muse Spark Open — thenewstack.io; Let's Data Science, "Meta Releases Muse Spark Open for Macs" — letsdatascience.com
  13. WOWTALE, "Tencent Opens the Weights of a 770B Model" (Aug 29, 2026) — en.wowtale.net; Developers Digest, "Tencent Hy4 Preview: 770B Open MoE" — developersdigest.tech
  14. Asia Economy, "LG AI Research Institute Releases Korea's Largest K-EXAONE 2.0" — asiae.co.kr; AI Weekly, "LG AI Research Ships K-EXAONE 2.0 as an Apache 750B MoE" — aiweekly.co
  15. SK Telecom, A.X K2 model card — huggingface.co/skt/A.X-K2; RCR Wireless, "SK Telecom A.X K2" — rcrwireless.com
  16. Ranks 11–20, from each lab's model card or release post: Z.ai GLM-5 (the-decoder.com); Huawei openPangu-Ultra-MoE-718B (arXiv 2505.04519, hiascend.com); Mistral AI, "Introducing Mistral 3" (mistral.ai); StepFun Step 5 Preview (marktechpost.com); Nvidia Nemotron 3 Ultra (thenewstack.io); Snowflake Arctic (huggingface.co/Snowflake); MiniMax-M1 (huggingface.co/MiniMaxAI); Baidu ERNIE 4.5 (technode.com); Arcee Trinity Large (huggingface.co/arcee-ai); AI21 Jamba 1.5 Large (huggingface.co/ai21labs)