The best open-weight AI models — November 2026, ranked by League Score.
The best open-weight AI models are the ones you can download, self-host and fine-tune — and in our database 30 of the 59 models we track qualify. This post ranks the scored ones by League Score, measures the gap to the closed frontier, prices the leader against the overall #1 on a real workload, and flags where a score rests on thin benchmark coverage.
By AI Pro League · Published
Best open-weight AI models, by our AI model rankings: DeepSeek V4.1 Flash leads the open field at 82.0, and it is not close to the closed frontier — the overall leader, GPT-6 Astra, sits at 99.8, a gap of 17.8 points. Of the 30 open-weights models we track, 21 carry a League Score (the rest have no recorded benchmark components yet, so they are unranked rather than scored zero). The full open table lives in the Open-Weight Division, and our picks page ranks the same field in best open-source AI models.
The best open-weight AI models, ranked
| # | Model | Maker | League Score | Coverage | Blended $/1M | Licence |
|---|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash | DeepSeek | 82.0 | 55% | $0.52 | MIT |
| 2 | DeepSeek V4.1 Flash 1017 | DeepSeek | 72.8 | 40% | $0.52 | MIT |
| 3 | Kimi K3 | Moonshot AI | 70.6 | 70% | $6.00 | Kimi K3 License |
| 4 | Kimi K2.6 | Moonshot AI | 69.7 | 35% | $1.71 | Modified MIT |
| 5 | Kimi K3.1 Pro | Moonshot AI | 69.4 | 40% | $4.00 | Kimi K3.1 Pro License |
| 6 | Qwen3.8-Max | Alibaba Qwen | 65.0 | 65% | $3.00 | Qwen3.8-Max License |
| 7 | DeepSeek V4 Pro (0813) | DeepSeek | 63.2 | 60% | $1.98 | MIT |
| 8 | Qwen3.5-397B-A17B | Alibaba Qwen | 60.7 | 35% | — | Apache-2.0 |
| 9 | Mistral Medium 3.5 | Mistral AI | 59.1 | 20% | $3.00 | Modified MIT |
| 10 | Llama 4 Maverick | Meta | 47.7 | 15% | — | Llama 4 Community License |
The shape of the field matters as much as the order. The median scored open-weight model sits at 41.3, against an overall median of 65.0 across all 45 scored models — the open field is top-heavy: a strong leading group, then a long tail. Two of the top three are DeepSeek V4.1 Flash variants, and the leader does it at a blended $0.52 per million tokens — the joint-cheapest rate in this top 10, shared with its 1017 variant, and far under the rest.
How far behind the closed frontier is the best open-weight model?
17.8 League points, on the numbers above — DeepSeek V4.1 Flash at 82.0 versus GPT-6 Astra at 99.8. Read that gap with the coverage column in hand: the leader’s score covers 55% of our five-benchmark set, the overall leader’s covers 55%. Neither is a full record, and a wider benchmark record could move either score — the honest statement is that on the benchmarks both models have actually been measured on, the closed frontier is ahead by a clear but not overwhelming margin.
What the best open-weight AI models cost — the worked example
Price the two leaders on one workload — 10M input tokens and 2M output tokens, at the rates in our AI model pricing index (DeepSeek V4.1 Flash: $0.30 / $1.20 per 1M, checked Oct 18, 2026 against DeepSeek’s official pricing page; GPT-6 Astra: $10.00 / $50.00):
| Model | Input cost | Output cost | Total |
|---|---|---|---|
| DeepSeek V4.1 Flash (open #1) | $3.00 | $2.40 | $5.40 |
| GPT-6 Astra (overall #1) | $100.00 | $100.00 | $200.00 |
The overall leader costs 37× as much on this workload — and for the open leader, the hosted API is only one option: the weights can be self-hosted, where the bill becomes hardware instead of tokens. That option, not the score gap, is why the open field keeps its audience. Our prompt caching pricing explainer covers the cached rates that change this maths on repeated-context work.
Coverage: why some open-weight scores deserve a second look
The deepest benchmark record in the open field belongs to Nemotron 3 Ultra 550B-A55B: 85% coverage — yet it ranks only #13 among open models at 32.3, because its recorded components are genuinely mid-field. Depth of record and height of score are different things, and this table keeps both visible: a high score on a thin record (several top-10 entries sit at 35% coverage or below) is a claim about the benchmarks that exist, not the ones that don’t. Our benchmark explainers — LiveBench, SWE-bench Verified, GPQA Diamond — set out what each component does and does not prove.
Which licence do the best open-weight AI models actually ship under?
“Open-weight” is not one licence. The 21 scored open models in our database ship under 12 different licences — Apache 2.0, Apache-2.0, Kimi K3 License, Kimi K3.1 Pro License and more — and the terms decide what self-hosting actually permits: some are permissive (MIT, Apache-2.0), others are community or ecosystem licences with conditions attached. The licence column in the table above is part of the ranking, not fine print: check it before a model’s score talks you into building on weights you may not be able to ship.