The Top 10 LLMs Right Now (October 2026), Including Open-Weight Models
The gap between the best proprietary models and the best downloadable ones is still real, but it is smaller than the headlines suggest, and it looks very different depending on what you pay per task. This is a snapshot of the ten strongest large language models as of October 10, 2026, ranked by the Artificial Analysis Intelligence Index, with open-weight models marked so you can see exactly where they land.
Each model is listed once, using its best configuration on the Artificial Analysis Intelligence Index, an independent composite of reasoning, coding and knowledge benchmarks. Scores move every time a model is updated or re-tested, so treat the numbers as a snapshot, not a verdict. "Open" means the weights can be downloaded; it says nothing about the license terms, which we cover below.
Because of the three-way tie at the bottom, the table shows eleven models. Just behind them, Kimi K3 from Moonshot scores 44 and is MIT-licensed.
Closed models still lead, by a clear margin. The best open-weight model scores 46 against 58 for the top proprietary one. That is a gap of about twelve points, though the best open models sit within a handful of points of Meta's and xAI's current flagships.
Value is a different story from raw score. On Artificial Analysis's own cost-per-task measure, GPT-6.1 Sol reaches a score of 52 at about $0.72 per task, while Claude Opus 5.5 at maximum effort reaches 58 at about $5.98. Gemini 4 Argon scores 53 at about $1.99. If you run a lot of requests, the cheaper tier is often the sensible default and the top model is a tool for hard problems.
Effort settings matter as much as model choice. The same model can score several points apart depending on how much reasoning budget it uses, and cost scales with it. Always compare models at comparable settings.
MiMo-V2.6-Pro (Xiaomi). The top open-weight model on this index. It is a sparse mixture-of-experts model with about 1.02 trillion total parameters and about 42 billion active per token, a 1 million token context window, and an MIT license, which is the most permissive of the big releases.
Qwen3.8 Max (Alibaba). Alibaba released the weights of this Max-class model in August 2026. Two cautions: the downloadable version is text-only and differs from the hosted API version, and it ships under a custom license with revenue and user thresholds rather than Apache 2.0, so read it before building a business on it.
GLM-5.3 (Z.ai). Strong at agentic coding. Its predecessor was MIT-licensed, but GLM-5.3 uses a new custom license aimed mainly at very large hosting providers. Its smaller sibling, GLM-5.3-Flash, is MIT-licensed and far easier to run.
Kimi K3 (Moonshot AI). About 2.8 trillion total parameters with roughly 104 billion active and a 1 million token context, MIT-licensed. A very capable agentic model, but it takes serious hardware to self-host.
For a single GPU or laptop. Google's Gemma 4 31B is released under Apache 2.0 and is one of the best options you can realistically run on one high-end GPU. It scored 39 on Artificial Analysis's index at its April launch, which was an earlier version of the index, so that number is not directly comparable to the table above. DeepSeek V4.1 Flash, which scores 39 on the current index, is a popular budget pick through its API.
- Best quality, budget secondary: Claude Opus 5.5 or Claude Sonnet 5.5.
- Best quality per dollar through an API: GPT-6.1 Sol or Gemini 4 Argon.
- Need the weights, with the cleanest license: MiMo-V2.6-Pro or Kimi K3 (both MIT).
- Need to run it locally on modest hardware: Gemma 4 31B (Apache 2.0).
- Large hosted product: check the license thresholds on Qwen3.8 Max and GLM-5.3 before committing.
- Benchmarks measure specific tasks. Your own workload may rank these models differently, so test with real prompts before choosing.
- Several details above, including parameter counts and license terms, come from secondary sources and the vendors' own announcements. Check each model's official model card and license file before you deploy.
- This field changes fast. A new release or re-test can reshuffle the order within weeks.
For more on this topic, see our earlier posts on [open-weight LLMs in 2026](https://directory.drveri.com/blog/open-weight-llms-in-2026-how-close-are-they-to-gpt-5-and [the most cost-efficient LLMs for coding](https://directory.drveri.com/blog/the-most-cost-efficient-llms-for-coding-in-2026-ranked and [running coding LLMs locally](https://directory.drveri.com/blog/best-home-computers-for-running-coding-llms-locally-in-2026
- Artificial Analysis Intelligence Index leaderboard: https://artificialanalysis.ai/leaderboards/models
- BenchLM open-source leaderboard: https://benchlm.ai/best/open-source
- Gemma 4 analysis (Artificial Analysis): https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know
- GLM-5.3 license coverage (The New Stack): https://thenewstack.io/zai-glm-weights-license/
- MiMo-V2.6-Pro (Artificial Analysis): https://artificialanalysis.ai/models/mimo-v2-6-pro
How This Ranking Works
Each model is listed once, using its best configuration on the Artificial Analysis Intelligence Index, an independent composite of reasoning, coding and knowledge benchmarks. Scores move every time a model is updated or re-tested, so treat the numbers as a snapshot, not a verdict. "Open" means the weights can be downloaded; it says nothing about the license terms, which we cover below.
The Top 10 (October 2026)
| Rank | Model | Maker | Index score | Weights |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | 58 | Closed |
| 2 | Claude Sonnet 5.5 | Anthropic | 56 | Closed |
| 3 (tie) | Claude Fable 5.1 | Anthropic | 53 | Closed |
| 3 (tie) | GPT-6 Astra | OpenAI | 53 | Closed |
| 3 (tie) | Gemini 4 Argon | 53 | Closed | |
| 6 | GPT-6.1 Sol | OpenAI | 52 | Closed |
| 7 | Muse Spark 1.3 | Meta | 48 | Closed (open weights promised, no date) |
| 8 (tie) | Grok 4.7 | SpaceXAI | 46 | Closed |
| 8 (tie) | MiMo-V2.6-Pro | Xiaomi | 46 | Open (MIT) |
| 10 (tie) | Qwen3.8 Max | Alibaba | 45 | Open (custom license) |
| 10 (tie) | GLM-5.3 | Z.ai | 45 | Open (custom license) |
Because of the three-way tie at the bottom, the table shows eleven models. Just behind them, Kimi K3 from Moonshot scores 44 and is MIT-licensed.
What Stands Out
Closed models still lead, by a clear margin. The best open-weight model scores 46 against 58 for the top proprietary one. That is a gap of about twelve points, though the best open models sit within a handful of points of Meta's and xAI's current flagships.
Value is a different story from raw score. On Artificial Analysis's own cost-per-task measure, GPT-6.1 Sol reaches a score of 52 at about $0.72 per task, while Claude Opus 5.5 at maximum effort reaches 58 at about $5.98. Gemini 4 Argon scores 53 at about $1.99. If you run a lot of requests, the cheaper tier is often the sensible default and the top model is a tool for hard problems.
Effort settings matter as much as model choice. The same model can score several points apart depending on how much reasoning budget it uses, and cost scales with it. Always compare models at comparable settings.
The Open-Weight Models Worth Knowing
MiMo-V2.6-Pro (Xiaomi). The top open-weight model on this index. It is a sparse mixture-of-experts model with about 1.02 trillion total parameters and about 42 billion active per token, a 1 million token context window, and an MIT license, which is the most permissive of the big releases.
Qwen3.8 Max (Alibaba). Alibaba released the weights of this Max-class model in August 2026. Two cautions: the downloadable version is text-only and differs from the hosted API version, and it ships under a custom license with revenue and user thresholds rather than Apache 2.0, so read it before building a business on it.
GLM-5.3 (Z.ai). Strong at agentic coding. Its predecessor was MIT-licensed, but GLM-5.3 uses a new custom license aimed mainly at very large hosting providers. Its smaller sibling, GLM-5.3-Flash, is MIT-licensed and far easier to run.
Kimi K3 (Moonshot AI). About 2.8 trillion total parameters with roughly 104 billion active and a 1 million token context, MIT-licensed. A very capable agentic model, but it takes serious hardware to self-host.
For a single GPU or laptop. Google's Gemma 4 31B is released under Apache 2.0 and is one of the best options you can realistically run on one high-end GPU. It scored 39 on Artificial Analysis's index at its April launch, which was an earlier version of the index, so that number is not directly comparable to the table above. DeepSeek V4.1 Flash, which scores 39 on the current index, is a popular budget pick through its API.
How to Choose
- Best quality, budget secondary: Claude Opus 5.5 or Claude Sonnet 5.5.
- Best quality per dollar through an API: GPT-6.1 Sol or Gemini 4 Argon.
- Need the weights, with the cleanest license: MiMo-V2.6-Pro or Kimi K3 (both MIT).
- Need to run it locally on modest hardware: Gemma 4 31B (Apache 2.0).
- Large hosted product: check the license thresholds on Qwen3.8 Max and GLM-5.3 before committing.
Caveats
- Benchmarks measure specific tasks. Your own workload may rank these models differently, so test with real prompts before choosing.
- Several details above, including parameter counts and license terms, come from secondary sources and the vendors' own announcements. Check each model's official model card and license file before you deploy.
- This field changes fast. A new release or re-test can reshuffle the order within weeks.
For more on this topic, see our earlier posts on [open-weight LLMs in 2026](https://directory.drveri.com/blog/open-weight-llms-in-2026-how-close-are-they-to-gpt-5-and [the most cost-efficient LLMs for coding](https://directory.drveri.com/blog/the-most-cost-efficient-llms-for-coding-in-2026-ranked and [running coding LLMs locally](https://directory.drveri.com/blog/best-home-computers-for-running-coding-llms-locally-in-2026
Sources
- Artificial Analysis Intelligence Index leaderboard: https://artificialanalysis.ai/leaderboards/models
- BenchLM open-source leaderboard: https://benchlm.ai/best/open-source
- Gemma 4 analysis (Artificial Analysis): https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know
- GLM-5.3 license coverage (The New Stack): https://thenewstack.io/zai-glm-weights-license/
- MiMo-V2.6-Pro (Artificial Analysis): https://artificialanalysis.ai/models/mimo-v2-6-pro