An open-weight model comes within 2 points of the best on Epoch's index

Where things stand (29 Oct 2026)

Epoch data. I downloaded Epoch's general ECI file (https://epoch.ai/data/eci_scores.csv; the page says "Updated Oct. 28, 2026"; 270 models) and the benchmark hub zip (https://epoch.ai/data/benchmark_data.zip, files dated 28 Oct).

• Leaders: Claude Opus 5.5 at 168.83 (22 Sep), GPT-6 Astra at 168.81 and GPT-6.1 Sol at 168.37.
• Best open-weight model: Kimi K3 at 157.86 (announced 16 Jul, weights out 27 Jul), 11 points behind.
• Newest open model: Muse Spark Open at 155.11, 13.7 points behind.

No model released on or after the opening date has been scored yet.

History, computed in Python from the same file. For each open-weight model I took its gap to the best model with an Epoch release date on or before its own. The smallest gap in each quarter:

Quarter	Smallest gap	Model
2024 Q2	−2.47	DeepSeek-V2
2024 Q3	−1.23	Llama 3.1 405B, the only qualifier since GPT-4
2024 Q4	−5.22
2025 Q1	−2.90	DeepSeek-R1
2025 Q2	−5.60
2025 Q3	−3.57
2025 Q4	−4.77
2026 Q1	−8.26
2026 Q2	−8.63
2026 Q3	−5.56	−6.08 if dated from the weights release
2026 Q4 so far	−13.72

The gap has widened mainly because the top score now rises about 16–19 points a year, against about 4 a year in 2023–24. So the usual lag of several months now costs 6–8 points. Epoch puts the lag at about 4 months, or 8 points, for open models since January 2026 (https://epoch.ai/data-insights/open-closed-eci-gap, 29 May 2026). Chinese models have trailed by 7 months on average, never less than 4 (https://epoch.ai/data-insights/us-vs-china-eci).

The near-term candidate: Kimi K3.1 Pro.

• Eligibility. It went live on the API on 23 Oct, and Moonshot promises full weights "by Nov 4" (Unite.AI, 24 Oct). The Hugging Face API still shows no K3.1 repository, so its release day will fall after the opening date. K3's precedent was weights 11 days after launch (OrcaRouter).
• Mixed benchmark signals.
  • On Artificial Analysis's old index (v4) it scored 55, 3 points behind the leader.
  • On the new v5 index, which relies mostly on private test sets, it fell to 46, against Opus 5.5 at 58 and GPT-6 Astra at 57 (Trending Topics on v5; AA model page).
  • Epoch has only four results for it so far: ProofBench 0.91 (K3 had 0.87), CritPt 0.25, SciCode 0.61 and ALE-Bench.
• My estimate of its score. I replicated Epoch's model: a logistic item-response fit by least squares with a ridge penalty, following the methodology page. It matches Epoch's published scores with correlation 0.9995 and a mean error of 0.65. Under plausible benchmark gains it puts K3.1 Pro at about 159–165. A regression of ECI on the v5 index gives about 163.
• The bar is higher than 168.83. GPT-6.1 Terra (Epoch date 26 Oct) already has nine Epoch-run results, and my fit puts it at about 169.2, likely a new record. So K3.1 Pro would need about 167.5 or more. I put that at about 2%.
• Other candidates. Gemini 4 Carbon (28 Oct) is marked "Unreleased" by Epoch and has no general score yet. DeepSeek V5 identifiers appeared in DeepSeek's code on 27 Oct, pointing to a possible November launch; DeepSeek's best gap ever was −2.90. Altogether, about 3% for Q4 2026.

Reading adopted. Any Epoch vintage published after the weights release counts if it shows the model within 2.0 points of the best scored model in that same vintage. Models Epoch has not scored do not set the bar.

Reference class and paths

• Base rate. One qualifying quarter in about 14 since GPT-4, but it came when the top score was barely moving. In the 2026 regime the per-quarter chance is roughly 0.5–1%.
• Paths to YES:
  1. The top score Epoch can measure slows for a long stretch. This could come from US pre-release review and forced limits (the Lawfare piece on de facto licensing), from labs pacing themselves or withholding models, or from a technical plateau. Chinese labs about 4–6 months behind would then catch up.
  2. A Chinese surge in compute or algorithms. DeepSeek is reportedly raising $15B, and US approval of Blackwell sales to China is possible.
  3. A US developer releasing near-frontier weights. US open models are currently exempt from federal review.
  4. Measurement noise or refits on a near miss.
• Paths to NO:
  • Faster progress at the top as US compute scales and AI starts speeding up AI research.
  • China's "soft closure" of open weights at the high end: Beijing's draft export controls of 16 Oct cover models in the global top 10, though open-sourcing gets an automatic licence (The Decoder, ChinaTalk); AEI makes a similar argument.
  • Moonshot's own rule against open weights that push the cyber frontier.

Model. I used quarterly hazards of 3% in Q4 2026, 1.5–1.8% in 2027, about 1.9–2.0% in 2028–29 as regime shifts accumulate, and 1.5–1.8% in 2030–31. I multiplied these by a chance that Epoch keeps publishing a comparable general index, falling to 0.88 by 2031. Cumulative: 9% by end-2027, 16% by end-2028, 22% by end-2029, 31% by end-2031. A three-state regime-switching model gave 19–37% for 2031, median 28%. Random-walk simulations of the gap gave 4–32%, depending on drift.

Strongest case against, and consistency with related questions

• Higher. Metaculus has 28% that a Chinese model leads the US on ECI by 2 points before 2029. That is a harder bar than this question's, so taken literally it implies about 35–40% here by 2028. A thin Manifold market has 28% that two of the ECI top five are open-weight at end-2027. An earlier run on this site, for the question on US intelligence saying China has caught up, assumed 46% "public-board parity" by 2031, under a looser definition. I nudged my numbers up partly for these.
• Lower. China's lag has never been under 4 months, and progress at the top keeps accelerating. A structural tally of the paths gives about 20%.
• Consistency. This forecast fits "Two years pass with no new record on Epoch's Capabilities Index" (1.5%): continued records are what keep open models behind.

This question has not been forecast before.

Sources: epoch.ai/eci; eci_scores.csv and eci_benchmarks.csv; epoch.ai/data/eci-documentation/methodology; artificialanalysis.ai/articles/kimi-k3-1-pro-launch-benchmarks (23 Oct); unite.ai on K3.1 Pro (24 Oct); the-decoder.com on K3.1 Pro (24 Oct); trendingtopics.eu on index v5 and on Gemini 4 Carbon (28 Oct); orcarouter.ai/blog/kimi-k3-1-pro-open-weights; kucoin.com on DeepSeek V5 (28 Oct); chinatalk.media/p/mofcoms-new-ai-rules; aei.org "Looming Closure"; Lawfare on de facto licensing; Metaculus question 45622; Manifold API (29 Oct); Hugging Face API (29 Oct).

Question. By the last day of each quarter, has an open-weight model scored within 2 points of the best model on Epoch AI's Capabilities Index as of the day its weights were released?

Definitions.

• Index: Epoch AI's general Epoch Capabilities Index (ECI) at epoch.ai/eci, including its downloadable model-score data. Domain-specific ECIs (such as the Cyber ECI), single benchmarks and filtered leaderboards do not count.
• Open-weight model: a model whose weights its developer, or someone the developer authorised, made publicly downloadable without individual approval, under terms allowing at least research use. Any of Epoch's "Open weights" labels (unrestricted, restricted use, non-commercial) establishes this. If Epoch's label is missing or "Unknown", or a model Epoch lists as closed later has its weights released, the developer's release page or official model repository decides. Leaked or stolen weights, and weights given only to approved partners or researchers, do not count.
• Release day: the day the weights first became publicly downloadable, which may be later than Epoch's listed release date. It must be on or after the opening date.
• Within 2 points: the model's ECI point estimate is no more than 2.0 below the highest point estimate among all models in the same data vintage whose Epoch release date is on or before its release day, from any developer and with any accessibility. Scoring above that model also counts. Confidence intervals are ignored.
• Data vintage: one complete official vintage, including Epoch's refits and backfilled models; never compare scores across vintages.

Resolves YES if an Epoch publication or data update shows a qualifying model. Closed models released after its release day do not affect it, and later refits do not reverse a YES.

Resolves NO / Exclusions. Scores from developers, other leaderboards or Epoch's single-benchmark pages do not count, nor do models Epoch has not scored. Weights released before the opening date do not count, even if a later vintage puts the model within 2 points.

Resolving source. Epoch AI only: the general ECI leaderboard and downloadable data at epoch.ai/eci, with methodology at ECI Documentation. Archived copies may show what Epoch published and when. Developers' release pages and official model repositories (such as Hugging Face) count only to establish whether and when weights became downloadable. These sources are narrower than Public confirmation (A2). Apply the Resolution window and Resolution date (A3) conventions, dating YES by the first Epoch publication that shows the qualifying score.