Intel / article

Kimi K2 6 The New Leading Open Weights Model

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published Apr 20, 2026

Context

Moonshot AI's Kimi K2.6 scores 54 on the Artificial Analysis Intelligence Index, the highest of any open weights model, whose weights anyone can download and run (artificialanalysis.ai). That places it fourth overall, behind only Anthropic, Google and OpenAI, which were then tied at 57; three days later GPT-5.5 broke that tie. The architecture did not change: like Kimi K2 Thinking and Kimi K2.5, it is a model with 1T total parameters, of which 32B are active for each token.

The biggest gain is on GDPval-AA, which Artificial Analysis calls its leading metric for general agentic performance. Models work through knowledge tasks such as preparing presentations and analysis, using code execution and web browsing in Stirrup, Artificial Analysis's open source agent harness. Kimi K2.6 reaches an Elo of 1520, up from 1309 for Kimi K2.5, and it maintains a 96% score on τ²-Bench Telecom, a test.

It also guesses less. Its rate, how often it fabricates an answer instead of abstaining when uncertain, fell to 39% from Kimi K2.5's 65%, close to Claude Opus 4.7 at 36% and MiniMax-M2.7 at 34%. The cost is verbosity: K2.6 used about 160M reasoning tokens to run the full index, significantly more than K2.5, slightly fewer than Claude Sonnet 4.6 at about 190M, and well above GPT-5.4 at about 110M. It accepts image and video input natively, keeps a 256K , and is served by Moonshot's API and by Novita, Baseten, Fireworks and Parasail.

Key takeaways · AI-distilled

Kimi K2.6 reached a GDPval-AA Elo of 1520, up from 1309 for Kimi K2.5, measured in Artificial Analysis's open-source Stirrup agent harness with code execution and web browsing. It holds 96% on τ²-Bench Telecom.
Its hallucination rate fell to 39% from Kimi K2.5's 65%, similar to Claude Opus 4.7 (36%) and MiniMax-M2.7 (34%), meaning it more often abstains when uncertain.
Kimi K2.6 used about 160M reasoning tokens to run the full index, slightly below Claude Sonnet 4.6 (~190M) but well above GPT 5.4 (~110M).
It keeps the 1T-total, 32B-active MoE design of Kimi K2 Thinking and K2.5, adds native image and video input with a 256K context, and is served by Moonshot plus Novita, Baseten, Fireworks and Parasail.

Terms in this piece · Glossary

mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source artificialanalysis.ai