AI · Open Weights

Open-Weight AI Frontier

The best model anyone can download, 5.6 points behind the closed frontier when it landed.

• LLaMA-65B’s 109.97 → Kimi K3’s 157.86
• Closed-frontier gap: 9.4 end-2024, 5.6 July 2026
• Downloadable, any licence — not open source
• A joint refit can rewrite this history

The record line takes a point each time a model with downloadable weights sets a new best score on Epoch AI’s Capabilities Index, dated to that model’s release, and runs from Meta’s LLaMA-65B at 109.97 in February 2023 to Moonshot’s Kimi K3 at 157.86 in July 2026. Epoch fits a single item-response model across more than fifty benchmarks, treating each as an exam question with its own difficulty, and reads off one capability number per model, a way of comparing models years apart as individual benchmarks saturate and are retired. The scale has no natural zero, so Epoch pins Claude 3.5 Sonnet at 130 and GPT-5 at 150, anchoring it on two closed models.

Downloadable weights first passed 130 with Microsoft’s Phi-4 in December 2024, under six months after Claude 3.5 Sonnet shipped, and first passed 150 with Moonshot’s Kimi K2.6 in April 2026, eight months after GPT-5. The closed frontier on the same index, derived the same way from the same file but not plotted here, ran 7.3 points ahead at the end of 2023, 9.4 at the end of 2024 as reasoning models arrived, and 10.2 at the end of 2025. On 16 July 2026 Kimi K3 closed to within 5.6 points of the standing closed record, Anthropic’s Claude Fable 5 from June. It is not the narrowest it has been: the gap was 1.23 points when Llama 3.1-405B landed in July 2024 and 2.9 when DeepSeek-R1 landed in January 2025.

The gap is a sawtooth, widening on every closed release and collapsing on every open one: three closed records in September 2026, Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5, reopened it to 11.0 points. Every figure here is the gap on its own date rather than today’s. Measured in months, an open-weight release in mid-2024 was matching what closed models had reached thirteen to fifteen months earlier, and since Llama 3.1-405B in July 2024 the lag has sat between one and eight months without narrowing further. Points and months disagree because the closed frontier does not advance evenly.

This is the most constructed curve on the site: ECI is a retrospective fit rather than a measurement, refitted jointly over every model and benchmark, so adding either shifts scores that already exist and can rewrite this line, changing a plotted value or which model held the record at all. Epoch publishes 90% bootstrap intervals for these scores, and the chart draws a point where the source has a range: two of these records are separated by 0.20, on intervals that almost entirely overlap. Epoch also warns that open-weight models appear to optimise for benchmarks more aggressively than proprietary ones, which would make this gap narrower on the page than it is in practice. Coverage is deliberate rather than exhaustive and needs at least four benchmark results per model, so a downloadable model Epoch never evaluated cannot appear.

Open-weight means downloadable, not open source: unrestricted, restricted-use and non-commercial licences all count, which is why Kimi K3 holds the record under a licence forbidding commercial use where GLM-5.2 before it did not.

Source: Epoch AI, Epoch Capabilities Index

Download the data: open_weight_frontier.csv · open_weight_frontier.json

All curves · The chart is drawn in your browser and needs JavaScript. Every number it plots is in the table above.