Kimi K2 — reviews, specs & pricing

Moonshot AI's trillion-parameter open MoE model for agentic tasks.

Summary

Kimi K2 is a 1T parameter (32B active) mixture-of-experts model trained for agentic tool-use and coding, released with open weights under a modified MIT license. It achieves state-of-the-art results among open models on agentic benchmarks. It supports a 128k context window.

Sample use case

Used for autonomous coding agents, tool-calling workflows, and research into large-scale MoE training. Popular among developers wanting an open GPT-4-class agentic model.

Specifications

  • Provider: moonshot ai
  • License: open
  • Parameters: 1000B
  • Context: 128k tokens
  • Input price: $0.6/M tok
  • Output price: $2.5/M tok
  • Released: 2025-07-11

Pros

  • Open weights at frontier scale
  • Strong agentic/tool-use performance
  • Low API cost

Cons

  • Massive compute needed to self-host
  • Newer with less track record
  • Chinese-lab model, data governance questions for some users

Average rating 4.3 from 4 community reviews on Reviuws.

Community reviews

Simon Willison's blog on Kimi K2

Rating: 4.0 / 5 — by Simon Willison (use case: open-weights experimentation)

Willison calls the Kimi K2 release colossal, detailing its trillion-parameter, 32B-active mixture-of-experts architecture and Muon optimiser training as a major open-weights milestone from a young lab.

Pros: Large but efficient MoE design; novel training approach documented in a paper.

Cons: 958GB footprint makes local hosting impractical; very new lab with a short track record.

VentureBeat on Kimi K2

Rating: 4.0 / 5 — by Emilia David (use case: coding and autonomous agents)

VentureBeat reports Kimi K2 challenging proprietary systems from OpenAI and Anthropic, with particularly strong coding and autonomous agent benchmarks, released free and open.

Pros: Beats GPT-4 on several benchmarks while free and open; strong at coding and agent tasks.

Cons: Trillion-parameter scale needs serious infrastructure; limited independent validation at the time.

VentureBeat on Kimi K2

Rating: 5.0 / 5 — by Michael Nuñez (use case: agentic reasoning and tool use)

VentureBeat reports Kimi K2 Thinking outperforming GPT-5 and Claude Sonnet 4.5 on key benchmarks, evidence that Chinese open-source labs are closing the gap with leading proprietary models.

Pros: Beats GPT-5 and Claude Sonnet 4.5 on cited benchmarks; leading free alternative.

Cons: Wins are task-specific rather than across the board; modified MIT licence isn't fully unrestricted.

the-decoder on Kimi K2

Rating: 4.0 / 5 — by Matthias Bastian (use case: agentic task automation)

Bastian positions Kimi K2 as the next open-weight milestone from China after DeepSeek, built to rival Claude Sonnet 4 and GPT-4.1, with agentic training and the MuonClip optimiser behind its benchmark showing.

Pros: Trained specifically for agentic capability; stable large-scale training approach.

Cons: Rivals rather than clearly beats top proprietary models; differentiation from DeepSeek questioned.

Frequently asked questions

What is Kimi K2?

Kimi K2 is a 1T parameter (32B active) mixture-of-experts model trained for agentic tool-use and coding, released with open weights under a modified MIT license. It achieves state-of-the-art results among open models on agentic benchmarks. It supports a 128k context window.

How much does Kimi K2 cost?

Kimi K2 costs $0.6 per million input tokens and $2.5 per million output tokens.

Is Kimi K2 open source?

Kimi K2 is released under the open license.