AI Change Tracker

Kimi K3

Lab
moonshot
Weights
open_weights
Released
2026-08-18
Modalities
text
API
yes
Local-runnable
no
Status
current

announcementmodel card

Scores

Hard reasoning

Value Source As of Reported by Notes
50 index artificial-analysis-intelligence-index ↗ 2026-09-07 aggregatorA Kimi K3 (max) — highest-ranked open-weights model, 38 open-weights of 95 models evaluated. Index is now v4.2 (10 evals); this is a rescaled ruler and is NOT comparable to the 60 recorded on 2026-08-31 under v4.1.1. Gap to the top model (Claude Fable 5.1 max, 57) reads 7 points on this scale.
60 index artificial-analysis-intelligence-index ↗ 2026-08-31 aggregatorA Kimi K3 (max) — highest-ranked open-weights model of 98 open-weights entries, 3 points off the top model. Index v4.1.1.

Human preference

Value Source As of Reported by Notes
1489 elo lmarena-text ↗ 2026-08-31 aggregatorA kimi-k3-max, #10 and the top open-weights entry on the text board; unchanged in value from the 2026-08-26 fetch, as-of refreshed.
1489 elo lmarena-text ↗ 2026-08-26 aggregatorA kimi-k3-max variant, tenth place on the 2026-08-26 board fetch; reported ~2.8T-parameter MoE — not locally runnable.

Signals

Events

4

Anthropic launches Claude Haiku 5.5, price-matched to GPT-6 Luna, and halves Sonnet 5.5 cache-read price

Anthropic released Claude Haiku 5.5, priced at $0.10/$0.50 per million input/output tokens under 100K tokens — matching OpenAI's GPT-6 Luna — rising to $0.50/$2.50 above that threshold. Anthropic says it costs about 75% less to run than Haiku 4.5 on average and is positioned as a cheap subagent paired with Opus 5.5 or Sonnet 5.5. The same day, Anthropic halved Sonnet 5.5's prompt-cache-read price from $0.20 to $0.10 per million tokens, added monthly API credits to Max and Team subscription plans, and built computer-use/browser-use toolsets into the Python and TypeScript SDKs. Independent testing by Artificial Analysis put Haiku 5.5 at an Intelligence Index of 43 at max effort, ahead of GLM-5.3 Flash (42), Gemini 3.8 Flash (41) and GPT-6 Luna (38), and comparable to Kimi K3 (44); Simon Willison noted Haiku 5.5 uses a less generous tokenizer, around 1.25x more tokens than Haiku 4.5 for the same prompt.

2026-10-07 Release Pricing change Benchmark result Introducing Claude Haiku 5.5
3

Interconnects briefs Congress: Chinese open-weight models now lead American open models by a wide AAII margin

In prepared remarks for a Congressional briefing on U.S.–China open-weight model competition, Nathan Lambert (Interconnects) said that as of September 14, 2026, the top three Chinese open-weight models — Z.ai's GLM-5.3 (Artificial Analysis Intelligence Index score 45), GLM-5.3-Flash (42), and Moonshot AI's Kimi K3 (44) — outscore the leading American open-weight models, Thinking Machines' Inkling and Inkling Small (26 each) and Nvidia's Nemotron 3 Ultra (23). He also cited Hugging Face download data showing China's cumulative open-weight download lead has grown to about 1.6B more downloads (roughly 3.2B China vs. 1.6B U.S. total) since he began tracking in August 2025, with China's Hugging Face download lead dating to July 2025 driven largely by Alibaba's Qwen models.

2026-09-21 Benchmark result Lab strategy The current balance of power in open models

← Back to Models