Model Cadence
Log in
NOTE · SEPTEMBER 15, 2026

Open weights have trailed the frontier by four to eight months since spring 2025

Every time an open-weights model set a new best score since January 2025, the closed frontier had already been there for between 1 and 8 months. Kimi K3 brought the gap to 4.4 months, the low end of the range, not a new regime.

Line chart of months behind the frontier for each new best open-weights model, January 2025 to July 2026, ranging between about 1 and 8 months.
Each point is a new best open-weights score on the Epoch Capabilities Index; its height is how long the closed frontier had already been at that level. Download this chart · Open it on Analytics
DATA AS OFSep 15, 2026
SOURCEEpoch AI Capabilities Index, CC BY 4.0, joined to release dates from Models.dev and lab announcements
PUBLISHEDSep 15, 2026

The question people actually ask about open weights is not "how far behind" but "is it closing". The chart answers the second one, and the answer so far is no.

What the chart measures

Take every open-weights model that set a new best score on the Epoch Capabilities Index since January 2025. For each one, find the first day any model, open or closed, had reached that score. The distance between the two dates is how far behind the open frontier was on the day it moved. It is the same idea as the frontier lag shown on every model page, applied only to the moments the open-weights record changed.

What it shows

  • The record moved twelve times in twenty months, so the open frontier is not standing still. DeepSeek moved it four times, Moonshot four, Alibaba twice, Z.ai twice.
  • The lag has stayed inside a band. Since spring 2025 the lowest point was 3.8 months (Qwen3-235B Thinking, July 2025) and the highest 8.0 months (GLM-5.1, April 2026). Before that, DeepSeek-R1 in January 2025 came within about a month of the frontier o1 had set five weeks earlier, and that remains the closest approach.
  • Kimi K3, released July 16, 2026, is at 4.4 months. It scores about what GPT-5.4 Pro scored in March 2026. That is the best figure since mid-2025, and it sits inside the band rather than below it.
  • The closed frontier keeps moving while the open one catches up. GPT-6 Astra took the overall lead on September 3, so the gap measured today is against a newer target than the one Kimi K3 launched against.

What it does not show

  • Scores are Epoch's current estimates, plotted at release dates. Epoch revises scores as it adds benchmark evidence, so a point can move after the fact; this chart is the snapshot on the data date.
  • "Open weights" follows Epoch's accessibility record for each model, not a licence review. A model with downloadable weights under a restrictive licence counts as open here.
  • The index is a composite of benchmarks. A model can be closer to the frontier on the one task you care about and further on another; the per-benchmark boards are where to check that.
  • Coverage is uneven. A model Epoch has not run does not appear, and that is a gap in measurement, not a score.

The chart above updates on Analytics as Epoch's data changes. The numbers in this note are fixed to September 15, 2026.

Every number above is read from the same data the site serves. Scores come from Epoch AI under CC BY 4.0 and can change as Epoch adds evidence; the chart is the snapshot on the data date. Quote it with the date.