MoonshotAI
Language model

Kimi K2 Thinking

Frontier-class reasoning at a fraction of the price.

4.5

At a glance

Provider
MoonshotAI
Type
Language model
Context window
262,144 tokens
Released
November 2025
Model ID
moonshotai/kimi-k2-thinking

Overview

Kimi K2 Thinking is Moonshot AI's reasoning-focused model, a trillion-parameter mixture-of-experts system (32B active) with a 256K context window that generates extensive chain-of-thought for hard math, coding and science problems. Released November 2025, it posts results that would have been frontier six months earlier — ~71% SWE-bench Verified, ~83% LiveCodeBench — at API prices several times cheaper than top Western models.

Moonshot is one of China's "AI tigers," and K2 Thinking is its statement that excellent open-weight reasoning models can be trained and shipped globally. It ships under a modified MIT licence through an OpenAI-compatible API, and the blunt takeaway from independent reviewers is simple: if your spend is dominated by an expensive frontier model and you haven't benchmarked Kimi, you may be overspending.

What it can do

  • Extensive chain-of-thought reasoning
  • 1T-parameter MoE with 32B active
  • 256K context window
  • Open weights, OpenAI-compatible API

Pricing: Very low for its capability — open weights available.

Benchmarks

SWE-bench Verified
~71%
LiveCodeBench
~83%
Price
Several× cheaper than frontier

Pros and cons

What works

  • Frontier-class reasoning at low cost
  • Strong agentic and coding scores
  • Open weights, easy to integrate

Trade-offs

  • Heavy chain-of-thought can add latency
  • Named alongside peers in 2026 distillation scrutiny

The verdict

Kimi K2 Thinking is one of 2026's clearest value stories — reasoning that rivals far pricier models, with open weights and an easy API. For math, coding and agentic work on a budget, it deserves a place in every benchmark shortlist.

Best for: Reasoning, math and coding at frontier quality on a budget.

Compare it

Questions

What is Kimi K2 Thinking best at?

Deep reasoning tasks — competitive math, complex coding and multi-step problem solving — where its extensive chain-of-thought pays off, all at a fraction of frontier pricing.

Is Kimi K2 open source?

Yes. It's released under a modified MIT licence with open weights and an OpenAI-compatible API, so it's easy to self-host or integrate.

Further reading

Put Kimi K2 Thinking at the table.

Ask Kimi K2 Thinking and other models the same question side by side, then keep the best answer.