TheRouter.aiTheRouter.ai
Back to Models

Kimi K2 Thinking

moonshotmoonshot/kimi-k2-thinking

Moonshot AI's deep reasoning model (1T total, 32B active MoE). Specialist for 200-300 step stable tool orchestration, long-horizon planning, and complex coding. Text-only.

Kimi K2 Thinking on Hugging Face ↗Kimi K2 Thinking announcement blog ↗Kimi K2 Thinking on OpenRouter ↗Moonshot AI platform docs ↗

Kimi K2 Thinking, released November 6, 2025 by Moonshot AI, is the thinking/reasoning variant of the open-weight Kimi K2 model family. It retains the same 1-trillion-parameter, 32-billion-active Mixture-of-Experts (MoE) architecture — 384 experts with 8 selected per token — but adds native chain-of-thought reasoning fused with real-time tool calling. Post-training uses quantization-aware training (QAT) at INT4 precision, shrinking the model footprint from ~1 TB to ~600 GB while maintaining benchmark competitiveness.

K2 Thinking's key differentiator is its ability to sustain 200–300 sequential tool calls within a single reasoning chain while maintaining coherent state — a capability that degrades rapidly in most models beyond 30–50 calls. It leads in agentic benchmarks like BrowseComp (60.2%) and HLE with tools (44.9%), outperforming GPT-5 in both. For TheRouter operators, it fills the premium reasoning+agent tier alongside moonshot/kimi-k2.6, offering an open-weight path for long-horizon autonomous research, multi-step code generation, and complex information retrieval pipelines.

Best for
  • • Long-horizon autonomous research: browse, search, cross-reference, and synthesise information across 100+ tool calls without losing coherence
  • • Agentic coding with self-correction: multi-step code generation, bug reproduction, log analysis, hypothesis testing, and iterative debugging via integrated tool use
  • • Deep reasoning on complex problems: mathematical proofs, multi-step logical deduction, academic research requiring synthesis across sources
  • • Self-hosted reasoning pipelines: open weights (Modified MIT License) allow fine-tuning on domain-specific trajectories and deployment on own GPU infrastructure
Reach for something else if
  • • Vision or multimodal inputs — K2 Thinking is text-only; route images/video to moonshot/kimi-k2.5 or moonshot/kimi-k2.6 which natively support multimodal input
  • • Latency-sensitive real-time chat where the extended thinking chain adds noticeable delay — keep moonshot/kimi-k2 (non-thinking) or moonshot/kimi-k2.6-instruct for faster turnarounds
  • • Simple Q&A where deep reasoning overhead is wasted — use the standard non-thinking K2 or K2.6 for cost efficiency
Context Length
256K
Max Output
64K
Input Priceper 1M tokens
$0.648/ 1M tokens
Output Priceper 1M tokens
$2.70/ 1M tokens

Modalities

text→text

Pricing Breakdown

TypeRate
Input$0.648 / 1M tokens
Output$2.70 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatreasoningstop

Specifications

Release date2025-11-06kimi-k2.org ↗verified
Architecture1T-total / 32B-active MoE; 384 experts, 8 selected per token; 61 layers; MLA attention; SwiGLU activation; INT4 QAT post-traininghuggingface.co ↗verified
Training cost$4.6 million (per CNBC, unverified independently)cnbc.com ↗to verify
Disk footprint (INT4)~600 GB (Hugging Face); vs ~1.03 TB for non-thinking K2huggingface.co ↗verified
LicenseModified MIT License — code and weightshuggingface.co ↗verified
Supported inference enginesvLLM, SGLang, KTransformers, NVIDIA NIM, mlx-lm (Apple Silicon)huggingface.co ↗verified
Training data cutoffNot publicly disclosedunknown

Benchmarks

BenchmarkDistributionScoreSource
Humanity's Last Exam (w/ tools)
HLE with tools; strong mode (8 parallel trajectories) reaches 51.0–51.3%. Compare: GPT-5 = 41.7%, DeepSeek-R1 = 20.3%
44.9%%lambda.ai ↗
BrowseComp
Autonomous web browsing and cross-source information synthesis. Compare: GPT-5 = 54.9%, Claude = 24.1%, DeepSeek = 40.1%
60.2%%lambda.ai ↗
SWE-bench Verified
Software engineering — real-world GitHub issues. Compare: GPT-5 = 74.9%, Claude = 77.2%, DeepSeek = 67.8%. Also CAISI NIST eval: 56.2%
71.3%%lambda.ai ↗
AIME 2025 (w/ Python)
AIME 2025 math competition with Python execution tool. CAISI NIST OTIS-AIME 2025 (no tools): 84.3%
99.1%%www.reddit.com ↗
GPQA-Diamond
Graduate-level Q&A (physics, chemistry, biology). Per CAISI NIST independent evaluation
83.8%%nist.gov ↗
MMLU-Pro
Massive Multitask Language Understanding (professional). Per CAISI NIST independent evaluation
89.3%%nist.gov ↗
CVE-Bench
Cybersecurity vulnerability exploitation. Per CAISI NIST independent evaluation
50.5%%nist.gov ↗
SMT 2025
Stanford Math Tournament 2025. Per CAISI NIST evaluation. Compare: GPT-5 = 91.8%, DeepSeek-R1-0528 = 87.6%
93.1%%nist.gov ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "moonshot/kimi-k2-thinking",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference →

Chat completion

Standard chat completion for K2 Thinking. The model will generate a thinking chain internally before producing its final answer. You can also stream the reasoning tokens.

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshot/kimi-k2-thinking",
    "messages": [
      {"role": "system", "content": "You are a helpful AI assistant."},
      {"role": "user", "content": "Design a distributed caching strategy for a global e-commerce platform."}
    ],
    "max_tokens": 8192,
    "temperature": 0.7
  }'

More from moonshot

moonshot/kimi-k2Base K2 non-thinking variant for faster, cheaper completions when deep reasoning isn't neededmoonshot/kimi-k2.6Latest K2.6 — multimodal, 1T-agent with 300 sub-agents, top-tier coding benchmarks (April 2026)moonshot/kimi-k2.6The live successor in Moonshot's K2 line.

Similar models

Cross-provider sibling models
deepseek/deepseek-r1
Leading open-weight reasoning model from DeepSeek — strong math and coding, but without native tool-calling fusion
qwen/qwq-plus
Alibaba Qwen's reasoning model — strong general reasoning, smaller footprint but fewer agentic tool-calling capabilities
qwen/qwq-32b
Smaller open-weight reasoning model (32B) from Qwen — suitable for local deployment with lighter hardware
openai/o3
OpenAI's premium reasoning model — highest pure-reasoning scores but closed-source and higher cost
anthropic/claude-sonnet-4.5
Claude's balanced reasoning model — strong extended thinking mode, alternative for tool-calling-heavy agentic workloads

News & changes

2025-11-06

Moonshot AI releases Kimi K2 Thinking — open-weight reasoning model with 200+ sequential tool calls

Moonshot AI announced Kimi K2 Thinking, the first open-weight reasoning model to natively fuse chain-of-thought with real-time tool calling. At $4.6M training cost, it achieved 44.9% on HLE (w/ tools) and 60.2% on BrowseComp — outperforming GPT-5 on agentic web search benchmarks. The INT4 QAT model (~600 GB) is available under the Modified MIT License.

re-authored by TheRouterkimi-k2.org ↗
2025-12-09

NIST CAISI evaluates Kimi K2 Thinking — strongest PRC-developed open-weight model at time of release

The U.S. NIST Center for AI Standards and Innovation independently evaluated K2 Thinking and found it to be the most capable PRC-developed open-weight model at release, though still behind top U.S. models on cyber and software engineering. The report confirmed K2 Thinking leads PRC models on SWE-bench (56.2%), GPQA (83.8%), and SMT 2025 (93.1%).

re-authored by TheRouternist.gov ↗
2025-11-08

Kimi K2 Thinking running on 2× M3 Ultra at 15 tok/s via mlx-lm

Apple MLX team member Awni Hannun demonstrated K2 Thinking running on two M3 Ultra Mac Studios (512 GB each) using pipeline parallelism in mlx-lm, achieving ~15 tokens/sec. The model's QAT INT4 format runs in native precision without quality loss, making local deployment of a 1T-parameter reasoning model practical for high-end Apple Silicon clusters.

re-authored by TheRoutersimonwillison.net ↗

Frequently asked

What's the difference between Kimi K2 (Instruct) and Kimi K2 Thinking?

Kimi K2 Instruct is a simple instruction-following (reflex-level) model optimised for fast responses without an explicit reasoning chain. K2 Thinking adds a full chain-of-thought reasoning phase before generating the answer and natively fuses thinking with tool calling — it can invoke tools mid-reasoning rather than waiting for thinking to finish. This makes K2 Thinking significantly more powerful for complex multi-step tasks, but also slower and more expensive per query. For simple Q&A, K2 Instruct is the more efficient choice.

kimi-k2.org ↗
Can I run Kimi K2 Thinking on consumer hardware?

Full precision (FP8) requires ~1.3 TB+ VRAM — roughly 12–16× H100 GPUs. The INT4 QAT variant (~600 GB) needs ~1.1 TB+ VRAM — 8× H200 is the recommended minimum. For Apple Silicon, the model has been demonstrated on 2× M3 Ultra (512 GB each) via mlx-lm at ~15 tok/s. Consumer GPUs (single 4090, 5090) cannot run the full model. However, the model is available via API through TheRouter and Moonshot's own platform without any hardware requirements.

lambda.ai ↗
How does K2 Thinking compare to DeepSeek-R1?

K2 Thinking outperforms DeepSeek-R1 across virtually all benchmarks. According to the Moonshot comparison table: HLE w/ tools: 44.9% vs 20.3%; BrowseComp: 60.2% vs 40.1%; SWE-bench Verified: 71.3% vs 67.8%. NIST CAISI's independent evaluation confirms K2 Thinking leads DeepSeek-R1-0528 on CVE-Bench (50.5 vs 36.0), GPQA (83.8 vs 81.3), and SMT 2025 (93.1 vs 87.6). K2 Thinking's key advantage is its fusion of reasoning with native tool calling — DeepSeek-R1 requires a separate loop for tool orchestration.

lambda.ai ↗
Is K2 Thinking fully open source?

Moonshot AI released K2 Thinking under a Modified MIT License, which permits use, modification, and distribution for most purposes but includes a restriction that the model may not be used to improve competing large language models. The weights, code, and training methodology are publicly available on Hugging Face. The license is not OSI-approved open source but is widely considered 'open-weight' and permits commercial use.

huggingface.co ↗
Fact ledger — every claim on this page traces here
sourceURLretrieved
Release datekimi-k2.org ↗2026-05-28verified
Architecturehuggingface.co ↗2026-05-28verified
Training costcnbc.com ↗2026-05-28to verify
Disk footprint (INT4)huggingface.co ↗2026-05-28verified
Licensehuggingface.co ↗2026-05-28verified
Supported inference engineshuggingface.co ↗2026-05-28verified
Training data cutoff——unknown
Humanity's Last Exam (w/ tools)lambda.ai ↗2026-05-28verified
BrowseComplambda.ai ↗2026-05-28verified
SWE-bench Verifiedlambda.ai ↗2026-05-28verified
AIME 2025 (w/ Python)www.reddit.com ↗2026-05-28to verify
GPQA-Diamondnist.gov ↗2026-05-28verified
MMLU-Pronist.gov ↗2026-05-28verified
CVE-Benchnist.gov ↗2026-05-28verified
SMT 2025nist.gov ↗2026-05-28verified
Moonshot AI releases Kimi K2 Thinking — open-weight reasoning model with 200+ sequential tool callskimi-k2.org ↗2026-05-28verified
NIST CAISI evaluates Kimi K2 Thinking — strongest PRC-developed open-weight model at time of releasenist.gov ↗2026-05-28verified
Kimi K2 Thinking running on 2× M3 Ultra at 15 tok/s via mlx-lmsimonwillison.net ↗2026-05-28verified
What's the difference between Kimi K2 (Instruct) and Kimi K2 Thinking?kimi-k2.org ↗2026-05-28to verify
Can I run Kimi K2 Thinking on consumer hardware?lambda.ai ↗2026-05-28to verify
How does K2 Thinking compare to DeepSeek-R1?lambda.ai ↗2026-05-28to verify
Is K2 Thinking fully open source?huggingface.co ↗2026-05-28to verify
Help & contact