Open-weights models will remain within one generation of the closed frontier through end of 2027.
confidence 0.7 · horizon 2027-12-31 · status open
Falsifiers
- No open-weights model within 10 AA index points of frontier for 6 consecutive months.
- Major open-weights lab converts to closed releases.
Evidence
- +1 DeepSeek open-sources a 'super-agent' harness prototype for multi-agent coordination — DeepSeek continuing to open-source frontier-adjacent tooling (not just model weights) extends its open-release cadence, broadening the open ecosystem's competitiveness beyond raw model benchmarks.
- +2 Mistral ships open weights for Mistral Large 4 ("Le Chonk"), a 1.05T-parameter model aimed at agentic coding and cybersecurity — The actual open-weight release of a trillion-parameter Mistral flagship, independently spec-confirmed by Artificial Analysis, extends the open-weights field to a second non-Chinese entrant this cycle, though the FrontierCode-#1 and win-rate claims remain lab-reported.
- -1 Artificial Analysis reweights its Intelligence Index to v5, widening the apparent open/closed capability gap — Artificial Analysis's reweighted v5 Intelligence Index now shows eight closed models beating the best open model, a wider gap than under the prior version — though this may partly reflect the benchmark's new weighting toward agentic-collaborator tasks rather than a real widening of open-model capability, so weight is capped low.
- +2 Moonshot launches Kimi K3.1 Pro, scoring near the closed frontier ahead of a promised Nov 4 open-weight release — A soon-to-be-open-weight model scoring 55 on Artificial Analysis's Intelligence Index, just 3 points behind the closed frontier leader's 58, is per Simon Willison the closest an open(-to-be) model has come to frontier parity on this independent benchmark, though weights aren't released yet.
- +1 DeepSeek reportedly raises $15B, one of China's largest funding rounds, days after US federal procurement ban — Fresh capital of this scale strengthens DeepSeek's ability to keep funding frontier-adjacent open-weights releases.
- +2 DeepSeek open-sources DeepInfer, its production inference stack, with large throughput/margin claims — DeepSeek open-sourcing its production inference stack, with lab-reported throughput/margin figures, extends its open-weights ecosystem lead and could push inference costs down market-wide, continuing DeepSeek's cadence of open releases within a generation of frontier.
- +2 Xiaomi's MiMo-V2.6-Pro-RL reported as top open model on Artificial Analysis's ranking — Independent measurement (Artificial Analysis) placing Xiaomi's MiMo-V2.6-Pro-RL as the top open model extends the open-weights field beyond DeepSeek/Qwen/GLM and further narrows the open/closed gap, though it rests on one independent ranking rather than corroborated multi-benchmark confirmation.
- +1 DeepSeek V4.1 Flash refreshed (1017 build) climbs to #2 open-weight on Artificial Analysis Intelligence Index — A continued independent-measurement benchmark refresh keeps an open-weight model competitive, extending the narrow open/closed gap pattern.
- -1 Mistral announces Mistral Large 4, scoring 38 on Artificial Analysis Intelligence Index; weights not yet released — A roughly 20-point Artificial Analysis Intelligence Index gap to the closed frontier for the leading non-Chinese open-weights lab's new flagship is mild counter-evidence, though weights are not yet released so it doesn't yet fully test the thesis.
- +1 Meta releases Muse Spark Open, an 816B-parameter open-weight model built to run on Mac — A new, independently-scored US open-weight entrant (39 on Artificial Analysis's Intelligence Index) extends the open-weights field, though it trails current open leaders GLM-5.3 and Kimi K3.
- +2 Reflection AI launches Beam, a 501B/23B-active open-weight model, positioning itself as a new US frontier-adjacent open lab — A new, well-capitalized US lab (Reflection AI) entering the open-weight frontier race with a 501B/23B-active model extends the open-weights field, though independent observers already place it behind current open SOTA (GLM 5.3, Kimi K3, DeepSeek V4.1 Flash), and its benchmark claims are lab-reported.
- +1 Anthropic's Frontier Red Team finds GLM-5.3 and an unreleased Claude Mythos model cross a cyber-offense capability threshold — GLM-5.3 (open) reaching a comparable capability level (4% full control-flow hijacks) to Anthropic's own unreleased Claude Mythos Preview (6%) on a cutting-edge offensive-cyber benchmark is further evidence of a narrow open/closed gap, this time on a safety-relevant capability rather than a standard benchmark.
- +1 Xiaomi releases MiMo-V2.6 open-weights models, hires ex-DeepSeek engineer who ran training in public — Xiaomi's reported MiMo-V2.6-Pro release, claimed via social-media roundup to be a new top open-weights model trained for about $3M, would extend the open field to a new lab and further narrow the open/closed gap if independently confirmed, but the claim is unverified beyond an aggregator repost.
- +1 Kimi-maker Moonshot AI targets $2B in annual revenue as K3 usage dips slightly — Moonshot's revenue ambitions plus continued heavy OpenRouter token volume for K3 signal it can keep funding frontier-adjacent open-weights releases even amid a reported usage dip.
- +1 Anthropic threat-intelligence report alleges distillation campaigns from Alibaba, Moonshot AI, DeepSeek; separately details rogue agents avoiding CAPTCHAs — Anthropic's lab-alleged report on persistent distillation campaigns by Alibaba, Moonshot AI, and DeepSeek suggests part of the open-weights field's narrow gap to frontier may come from distilling closed-model outputs rather than independent capability gains; unverified beyond Anthropic's own claim.
- +1 DeepSeek ships V4.1-Flash with a novel prefill/decode-split MoE architecture, retiring old Flash naming — A novel efficient MoE architecture continues DeepSeek's open-weights research cadence, though the curator notes it trails other open models on published benchmark headlines.
- +1 Mistral raises €3B Series D at €21B valuation, led by Samsung, Scaleup Europe, PSG Equity — A large funding round for the leading non-Chinese open-weights lab strengthens its ability to keep pace under its explicit open-weights commitment, though capital alone doesn't move the capability gap.
- +1 DeepSeek opens weights for DeepSeek-V4-Flash-Vision-Exp, adding vision to the V4 Flash line — An open, vision-capable DeepSeek V4 variant broadens open-weights feature parity with closed frontier multi-modal models, though the release is unverified independently.
- +1 Z.ai open-weights the full GLM-5.3 model (744B/40B active), positioned for agentic coding and cyber defense — A second Z.ai open-weights release (full GLM-5.3, 744B/40B active) within the same week extends the open frontier, though the claims are curator-relayed and unverified, capping the weight.
- +1 Tencent releases Hunyuan Hy4-preview, a 770B open-weights MoE claimed to lead SWE-bench Pro and jump on WebDev arena — A fourth Chinese lab (Tencent) reportedly fielding a top-tier open MoE would broaden the open-weights field beyond DeepSeek/Qwen/Z.ai, but claims are social-media-relayed and unverified, capping the weight.
- +1 Mystery model 'Ox Alpha' confirmed as GLM-5.3-Flash; independent quantization and serving benchmarks emerge — Independent quantization and serving benchmarks (Unsloth, Baseten, Databricks, Together AI) corroborate GLM-5.3-Flash's price-performance beyond the initial lab announcement, reinforcing the narrow open/closed gap.
- +1 Z.ai confirmed as the lab behind Ox Alpha, a mysterious open model reportedly topping leaderboards — A reportedly leaderboard-topping new open model would extend the open-weights field beyond DeepSeek/Qwen, but the claim is unverified and weights are not yet released, so weight is capped low.
- +2 Z.ai formally launches GLM-5.3-Flash, revealing it as the previously teased 'Ox Alpha' — An MIT-licensed 320B/18B-active model that independent measurement flagged for strong intelligence-per-dollar extends the open-weights frontier beyond DeepSeek/Qwen, though the Opus-4.8-parity coding claim is lab-reported and unverified.
- +1 Qwen releases Qwen3.8-Flash-Next as an early preview of Qwen4's architecture — Qwen open-sourcing an early Qwen4-architecture preview broadens the open field even before independent benchmarks exist.
- +2 DeepSeek releases V4 open weights under MIT license — Smallest open/closed gap on record
- +1 Mistral raises €4B, commits to keeping flagship weights open — €4B round tied to an explicit open-weights commitment through 2027.
Confidence history
- 2026-04-01 → 0.6 — Initial
- 2026-08-24 → 0.7 — DeepSeek V4 within 5 AA points