Top 10 Open-Weight LLMs in 2026 (Ranked by Arena Elo)
These are the ten highest-rated open-weight large language models by LMArena Chatbot Arena Elo, restricted to models whose weights you can actually download. Eight of the ten ship under genuinely open MIT or Apache 2.0 licenses. We flag which models carry restrictive custom licenses, and we name the source behind every score so you can verify it.
How We Ranked These Models
Ranked by LMArena Chatbot Arena Elo among models with publicly downloadable weights, as of Feb-Mar 2026 (LMArena, Onyx AI, LLM-Stats, and Iternal). Open weight is not the same as open source: license restrictiveness is flagged separately for each model.
Scores shift continuously as more head-to-head votes accumulate and new models enter the arena. Treat this order as directional, not absolute. Different trackers report slightly different figures for the same model, so we present a representative number and name the tracker it came from.
Arena Elo measures how often human voters prefer one model's response over another in blind head-to-head comparisons. It captures real-world helpfulness rather than a single static benchmark, which is why it is a useful primary sort for a general-purpose ranking. We pair it with one headline benchmark per model (SWE-bench Verified, MMLU-Pro, LiveCodeBench, GPQA Diamond, or HumanEval) so you can see where each model's measured strengths lie.
We restricted the list to models whose weights are genuinely downloadable. We then separated the leaders by license type. Models under MIT or Apache 2.0 are fully open. Models under custom vendor licenses, such as Meta's Llama Community License and Google's Gemma License, are open weight but carry usage restrictions, so we flag them in their own section rather than mixing them into the headline ten.
The Full Ranking
All ten ranked open-weight models at a glance, ordered by Arena Elo. Click any column header to sort, for example by Elo or by organization. Every score traces to the independent trackers named in the methodology above.
| # ↕ | Model ↕ | Organization ↕ | License ↕ | Arena Elo ↕ | Headline Benchmark |
|---|---|---|---|---|---|
| 1 | GLM-5 | Zhipu AI | MIT | 1454 | SWE-bench Verified 77.8 |
| 2 | Qwen 3.5 (397B) | Alibaba | Apache 2.0 | 1450 | MMLU-Pro 87.8 |
| 3 | GLM-4.7 | Zhipu AI | MIT | 1441 | SWE-bench Verified 73.8 |
| 4 | Kimi K2.5 | Moonshot AI | MIT | 1438 | SWE-bench Verified 76.8 |
| 5 | DeepSeek V3.2 | DeepSeek | MIT | 1423 | MMLU-Pro 85.0 |
| 6 | Qwen3-235B-A22B | Alibaba | Apache 2.0 | 1423 | GPQA Diamond 71.1 |
| 7 | Mistral Large 3 | Mistral | Apache 2.0 | 1416 | HumanEval 92.0 |
| 8 | MiniMax M2.5 | MiniMax | Apache 2.0 | 1404 | SWE-bench Verified 80.2 |
| 9 | DeepSeek R1 | DeepSeek | MIT | 1398 | MMLU-Pro 84.0 |
| 10 | MiMo-V2-Flash | Xiaomi | MIT | 1393 | MMLU-Pro 84.9 |
Elo figures are a representative Feb-Mar 2026 snapshot from LMArena, Onyx AI, LLM-Stats, and Iternal. Benchmark scores are the headline metric reported for each model on those trackers. Parameter counts: GLM-5 744B, Qwen 3.5 397B, GLM-4.7 355B, Kimi K2.5 1T, DeepSeek V3.2 685B, Qwen3-235B-A22B 235B, Mistral Large 3 675B, MiniMax M2.5 230B, DeepSeek R1 671B, MiMo-V2-Flash 309B.
Strong Models With Restrictive Licenses
These models post competitive Arena Elo scores and their weights are downloadable, but they ship under custom licenses rather than MIT or Apache 2.0. We keep them out of the headline ten so readers do not assume they carry the same freedoms. They are still worth knowing about, particularly Llama for its large fine-tuning ecosystem.
Open Weight Is Not the Same as Open Source
These two phrases get used interchangeably, but they mean different things, and the difference affects what you are legally allowed to do.
- Open weight means the trained model parameters are published and you can download them. You can run the model on your own hardware, fine-tune it, and inspect its behavior. Every model on this page is open weight.
- Open source is a stricter standard tied to a license that grants broad rights to use, modify, and redistribute without unusual restrictions. MIT and Apache 2.0 meet this bar. Custom vendor licenses, such as the Llama Community License and the Gemma License, generally do not.
Why it matters: a model can be open weight while still limiting commercial scale, restricting how outputs may be used, or carving out certain regions or use cases. Eight of the ten ranked models above use MIT or Apache 2.0, which is why they make the headline list. The restrictive-license models are flagged separately precisely so you can tell the difference at a glance. Before you build a product on any of these, read the actual license text rather than assuming open weight implies no strings attached.
Newer Models Without Public Elo Yet
Several newer open-weight releases had shipped at the time these sources were compiled but did not yet have a public Arena Elo. We do not assign scores to models that lack independent leaderboard data, so these are noted as pending rather than ranked:
- DeepSeek V4 (Pro and Flash): Released, but no public Arena Elo recorded in these sources yet.
- GLM-5.1: A newer revision of the top-ranked GLM family, not yet tracked on the arena.
- Kimi K2.6: A successor to Kimi K2.5, also awaiting public Elo data.
When these models accumulate enough head-to-head votes to register a stable Arena Elo, the standings above will change. That is the nature of a live leaderboard. We would rather mark a model as pending than publish a number we cannot trace to a source.
Frequently Asked Questions
What is the highest-ranked open-weight LLM in 2026?
As of the Feb-Mar 2026 snapshot, GLM-5 from Zhipu AI holds the top Arena Elo among open-weight models at roughly 1454 on independent trackers, narrowly ahead of Alibaba's Qwen 3.5. GLM-5 is released under the MIT license, which is fully open. Rankings shift as new models are evaluated, so treat the order as directional.
Is open weight the same as open source?
No. Open weight means the trained model weights are publicly downloadable. Open source is a stricter standard tied to a permissive license. Most models on this list use MIT or Apache 2.0, which are genuinely open. Others, such as Meta Llama and Google Gemma, ship under custom licenses with restrictions, so they are open weight but not open source in the strict sense.
Why are Llama 4 and Gemma flagged separately?
Both ship under restrictive custom licenses rather than MIT or Apache 2.0. The Llama Community License imposes a 700 million monthly active user cap, a clause against training competing models, and an EU multimodal carve-out. The Gemma License is a custom acceptable-use agreement. We separate them so readers do not assume they carry the same freedoms as the MIT and Apache-licensed leaders.
How current is this Arena Elo ranking?
The Elo figures reflect a Feb-Mar 2026 snapshot drawn from LMArena, Onyx AI, LLM-Stats, and Iternal. Arena Elo moves continuously as more head-to-head votes accumulate and as new models enter. Different trackers report slightly different numbers for the same model. Verify current standings on LMArena before making a decision.
Why are DeepSeek V4 and GLM-5.1 not on the list?
Newer releases such as DeepSeek V4, GLM-5.1, and Kimi K2.6 had shipped but did not yet have a public Arena Elo at the time these sources were compiled. We do not assign scores to models that lack independent leaderboard data, so they are noted as pending rather than ranked.
Video Resources
Go Deeper
Resources from across Tech Jacks Solutions
FREEAI Governance Charter
Establish your organization's AI principles in one document
FREEAI Risk Management Template
Identify, assess, and mitigate AI deployment risks
Prompt Engineering Library
Prompting techniques that get better results from any model
AI Glossary
Definitions for terms like open weight, Elo, and quantization