Moonshot AI Launches Kimi K3 with 2.8 Trillion Parameters
Simon Willison · rss · 2026-07-17
Moonshot AI has released Kimi K3, claiming it as their "most powerful model to date." It features 2.8 trillion parameters, is available via the official website and API, and its open weights are slated for release around July 27, 2026.
Key self-reported and third-party benchmarks include:
- In Artificial Analysis's long-horizon knowledge work evaluation, K3 achieved an overall Elo of 1547, outscoring Kimi K2.6 by 732 points and trailing only Claude Fable 5.
- Cost per task is roughly $0.94, closely matching GPT-5.6 Sol and about half that of Opus 4.8.
- Output token usage dropped by 21% compared to K2.6.
- K3 also surged to a leading position on Arena.ai's Frontend Code leaderboard.
Pricing is set at $3 per million input tokens and $15 per million output tokens, aligning closely with Anthropic's Claude Sonnet series, making it one of the priciest models from a Chinese lab to date.
The article's latter half uses the classic "pelican riding a bicycle" test to discuss benchmark limitations: while briefly relevant in the past, it no longer reflects a model's core capabilities, particularly in agentic tool calling and tool reliability during extended conversations.
More from Models
- Sol61 flunks instruction-following: skips mandated research even when explicitly told — lucasmeijer · 2026-10-26
- Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
- Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
- Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26
- supermemory unveils learner-1, claiming continual learning for any model — julianweisser · 2026-10-26
- Anthropic says Opus 5.5 cuts cost per task by ~40% vs Opus 5 — taufiqintech · 2026-10-26