AI CRITIQUE

Açık renkli bir zeminde yan yana iki ahşap saksı duruyor; soldaki küçük yapraklı bir filiz, sağdaki ise çiçek açmış tam boy bir ağaç barındırıyor. — Temsilî görsel — FLUX.1-dev ile bilgisayarımda üretildi; haberin gerçek fotoğrafı değil.Illustrative image — generated with FLUX.1-dev on my computer; not a real photograph of the story.

Mozilla: Open-source AI models lag 4.4 months behind

Summary

Mozilla's September 2026 'State of Open-Source AI' report numerically highlights the 4.4-month performance gap between open-weight and closed-source models. The report is based on Epoch AI's ECI index, METR's time horizon metric, and a survey of 1,400 developers. The top open model, Moonshot's Kimi K3, scored 157 on the ECI, while Anthropic's Claude Fable 5/Opus 5 model scored 162. In contrast, open models operate at approximately 30% lower cost per API call.

Why it matters

The report provides a concrete, multi-source benchmark showing how far open-weight AI has come and how far behind it remains compared to proprietary leaders. This directly impacts procurement decisions, regulatory discussions regarding model concentration, and the strategic calculations of companies building their own inference infrastructure. The 4.4-month gap is narrowing; open models double their task horizons every 3.9 months, compared to 5.5 months for closed models. While this indicates the open ecosystem is closing in, it is not fast enough to eliminate the autonomy and reliability advantages currently held by frontier labs.

Highlights

Kimi K3 (Moonshot) is currently the leader among open-weight models with an ECI score of 157. Compared to Anthropic's 162 points, this difference equates to approximately four months of R&D time.
Open models operate at approximately 30% lower cost per API call. Meanwhile, they experience only a 2.1-point loss in Artificial Analysis's price-performance index.
The 'sawtooth edge' expression indicates that open models are actually leading in web interface coding and have reached parity in agent browsing, but are lagging behind in specialized desktop tasks and long-form text analysis.

Critical take

The 4.4-month figure is a snapshot dependent on which composite index is trusted (ECI or METR horizon) and a limited number of named models. Therefore, it carries the risk of being over-interpreted as a stable structural chasm rather than a variable, task-dependent gap. Mozilla itself notes that open models may be over-optimized for specific benchmarks and that reliability drops sharply in autonomous tasks exceeding approximately 16 hours. These caveats suggest that the headline figure is less definitive than it appears.

Bu kritik, yerel yapay zeka modeli Qwen3.8-27B tarafından yazılmıştır. Haberi ben seçtim; kritik metnini yayımlanmadan önce tek tek denetlemiyorum.This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.

Back to the issue