Kimi K2 Thinking
200-Tool-Call Agent
Moonshot's K2 model trained to think while it acts. Most reasoning models think THEN call a tool. K2 Thinking interleaves reasoning and tool calls — chaining 200–300 actions per task without losing the plot. Released with open weights and a reported $4.6M training bill, it shocked the field by topping benchmarks at a fraction of Claude / GPT-5 cost.
Intelligence
Good
Speed
Slow
41 tok/s output
Cost
Moderate
$0.50 in / $2.85 out
Context
256K
Up to 256,000 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on
Artificial Analysis. Cost buckets are fixed dollar
thresholds keyed off output-token price ($/M out).
Intelligence
- Top 1%≤ 1%
- Top 5%≤ 5%
- Top 10%≤ 10%
- Good≤ 25%
- Medium≤ 50%
- Below avg> 50%
Speed
- Top 1%≥ 345 tok/s
- Top 5%≥ 237 tok/s
- Top 10%≥ 196 tok/s
- Good≥ 146 tok/s
- Medium≥ 90 tok/s
- Slow< 90 tok/s
Cost
- Freeopen weights · self-host
- Low< $1 / M out
- Moderate$1–5 / M out
- High≥ $5 / M out
Why it matters
Made long-horizon, tool-using agents accessible without an API contract. Proved that test-time compute scaling — chaining hundreds of tool calls — beats pure parameter scaling on many tasks.
Core Capabilities
Long Documents
Handles entire codebases, books, and multi-doc RAG.
Generative
Produces images, video, audio, or other media.
Agent Workflows
Built for tool use and autonomous tasks.
Reasoning
Solves complex math, logic, and planning tasks.
Context Window
256k tokens
≈ 197+ pages
4k Chat 聊天
32k Long docs 长文档
128k Books 整本书
400k Multi-doc 多文档
1M Codebase 整个代码库
10M
256k
Availability
API
Available
Product / App
Available
Open Source
Released
Enterprise
Contact sales
Pricing Model
Free / self-host
Open weights — pay only for compute.
Self-host Capability / Performance
Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).
Lower 20% Upper 80% This model
Context / memory
Context window size · log-scaled
6.0
9.0
7.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this
Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.
What it feels like
- 1T-parameter MoE / 32B active per token — Moonshot's open-weights debut, modified MIT license
- 65.8% on SWE-bench Verified single-attempt — outperforms every model tested except Claude Sonnet 4 at release
- 53.7% on LiveCodeBench v6 — strong open-source coding tier
Best use cases
- Self-hosted agent platforms where API models can't go (regulated, private cloud)
- Cost-sensitive frontier-tier inference via budget providers
- Coding agents that need long context + open weights for fine-tuning
Tools to try
Not ideal for
- Edge / single-GPU deployments — 1T MoE still demands multi-node serving
- Multimodal tasks (text-only at this generation; vision lives in Kimi-VL)