spécialisé deliberate modèle de raisonnement trained avec reinforcement learning pour slow recherche, multi-hour plan formation, et mathématiques proofs.
Technical Specifications
- Provider
- Moonshot AI
- Category
- reasoning
- Release Date
- 2025-11
- Batch API Discount
- Standard rate
- Consumer Subscription
- Free / Self-Hosted
- Reasoning Effort Control
- Configurable (Low / Med / High)
- Structured Outputs
- Yes
- Output Modalities
- text
- Official Documentation
- Visit Official Docs ↗
Data Privacy, Licensing & Deployment Governance
Official Provider Release Intelligence, API Quotas & Agentic Specs
- Official API Model ID (SDK String)
- kimi-k2-thinking
- Supported SDK & API Protocol
- OpenAI-Compatible REST / vLLM / SGLang / HuggingFace TGI
- Lifecycle & Deprecation Status
- Active GA (Guaranteed support through 2027+)
- Production Rate Limits (RPM & TPM)
- Unlimited (Hardware-bound on self-hosted GPU) / 10,000 RPM Cloud · Unlimited local throughput / 10M+ TPM Managed Cloud
- Long-Context Pricing Tier (>128K / >200K)
- Free across full context window (only KV-cache VRAM scales)
- Prompt Cache Write & TTL Policy
- Automatic prefix KV-caching in vLLM / SGLang (0 cost)
- Computer Use & Browser GUI Agent
- API / CLI Tool-Use & DOM Agent ready (no raw pixel desktop click loop)
- MCP & Parallel Tool Calling
- Native MCP Client/Server ready · Parallel function calling · 100% Strict JSON Schema
- Web Search Grounding & Code Sandbox
- Supported via external Tavily/Exa Search & E2B/Docker Sandbox tools
- Fine-Tuning & Distillation Support
- Full Weights SFT, LoRA, QLoRA, DPO & GRPO RL (Unsloth / Axolotl / Llama-Factory)
- Multimodal Tokenization Billing
- Text BPE (~4 chars/token)
- Weights Precision & Quantization
- BF16 / FP16, FP8 Dynamic, INT4 AWQ / GPTQ, GGUF (Q4_K_M / Q8_0)
- Minimum GPU VRAM Requirement
- 620 GB VRAM (4-bit Quantized) · 2150 GB VRAM (Native BF16)
- Supported Inference Engines
- vLLM, SGLang, Ollama, llama.cpp, TensorRT-LLM, HuggingFace TGI
- System Card & Safety Framework
- Open-Weights System Card + Customizable Guardrail Layer (Llama Guard / ShieldGemma)
- Enterprise Compliance (SOC2 / HIPAA / ZDR)
- 100% Air-Gapped On-Premises / Private VPC · Full HIPAA, GDPR & SOC2 Control
Kimi K2 Thinking Cost, Context, Hardware & API Integration
Open-weights or free-tier model: $0 API license cost when self-hosted.
Same Monthly Workload Cost Across Popular Reference Models
Live Prompt Paste Tokenizer & Cost Counter
Paste real system instructions, JSON schemas, or code to measure BPE token counts across tokenizers.
Hybrid Prompt Routing Cascade Simulator
Route routine queries to a fast lightweight model ($0.15/1M) and escalate only complex reasoning turns to Kimi K2 Thinking:
Benchmark Scores & 6-Axis Capability Radar
- SWE-bench Verified
- 50%
- GPQA Diamond
- 70%
- MMLU-Pro
- 82.5%
- HumanEval
- 93%
- AIME
- 82.5%
- Chatbot Arena Elo
- 1375pts
Strengths
- Exhaustive étape par étape mathématique raisonnement et vérification boucles
- Spacious 262K fenêtre de contexte accommodating dense scientifique whitepapers
- à poids ouverts version avec unrestricted académique et commercial utility
Limitations & Trade-offs
- étendu raisonnement chains increase response latence noticeably
- Overly verbose sur simple conversationnel requests
Editorial Assessment
un extraordinary ouvert deliberate reasoner conçu pour complexe problem-solving.
Best for: formel mathématique proofs, algorithmic architecture, et comprehensive scientifique recherche
Compare Kimi K2 Thinking
Frequently Asked Questions about Kimi K2 Thinking
How much does the Kimi K2 Thinking API cost per 1M tokens?+
Kimi K2 Thinking by Moonshot AI is priced at $0 per 1M input tokens and $0 per 1M output tokens. With prompt caching enabled, cached input tokens cost $0 per 1M tokens, and asynchronous Batch API requests receive a 0% discount.
Should I use a $20/month subscription or the Kimi K2 Thinking pay-as-you-go API?+
A flat $20/month consumer subscription is predictable for daily interactive chat and manual coding. However, if you are building software, running automated workflows, or sending low-to-moderate volume with prompt caching enabled, the pay-as-you-go API is typically far more cost-effective and offers higher rate limits without manual UI caps.
How many words and pages fit in Kimi K2 Thinking's context window?+
Kimi K2 Thinking features a 262K tokens context window, which equals approximately 196,608 English words or ~524 standard single-spaced pages in a single prompt. Its maximum output generation per turn is 32,000 tokens.
Can I run Kimi K2 Thinking locally on my own GPU (Open Weights)?+
Yes, Kimi K2 Thinking (1T total (32B active MoE)) releases open weights that can be self-hosted locally or on private cloud GPUs using vLLM, Ollama, SGLang, or llama.cpp.
Does Moonshot AI train on data sent to the Kimi K2 Thinking API, and is commercial use allowed?+
100% Isolated when self-hosted on private GPUs or VPC. License tier: Open-Weights Research & Commercial License. Full commercial ownership of generated outputs. Supported deployment channels include: Self-Hosted GPU (vLLM / Ollama), Private VPC, or Managed Serverless API.
How does Kimi K2 Thinking perform on coding and reasoning benchmarks?+
Kimi K2 Thinking achieves a composite Intelligence Index of 85/100, scoring 50% on SWE-bench Verified, 70% on GPQA Diamond, 82.5% on MMLU-Pro, and 1375 pts on Chatbot Arena Elo.
What is Kimi K2 Thinking best used for?+
Kimi K2 Thinking is best suited for: formel mathématique proofs, algorithmic architecture, et comprehensive scientifique recherche. Assessment: un extraordinary ouvert deliberate reasoner conçu pour complexe problem-solving.
Moonshot AI Family Lineage, Routing Tiers & Upgrade Path
Step down to a lighter sibling for high-volume routing or step up to Moonshot AI's flagship reasoning tier:
Kimi K3
Kimi K2.7 Code
Kimi K2.6
Kimi K2.5
Kimi K2 Thinking Turbo
Kimi K2
AI Glossary & Architecture Concepts
View full glossary →Key machine learning concepts and benchmarks relevant to Kimi K2 Thinking:
L'unité textuelle de base traitée par un modèle linguistique, représentant généralement 3 à 4 caractères ou environ 0,75 mot. Par exemple, une phrase courante est découpée en plusieurs jetons.
Une suite de tests standardisés permettant d'évaluer et de comparer objectivement les performances des modèles d'IA sur des tâches ciblées (code avec SWE-bench, mathématiques avec AIME, etc.).
Le nombre maximal de jetons qu'un modèle peut traiter simultanément (entrée et sortie comprises). Une fenêtre de 1 000 000 jetons permet d'analyser environ 750 000 mots en une seule requête.
Un modèle d'IA entraîné par apprentissage par renforcement pour allouer du temps de calcul à la réflexion logique avant d'émettre une solution, idéal pour les problèmes scientifiques et le code complexe.
Une méthode de raisonnement pas-à-pas où le modèle décompose le problème en étapes logiques intermédiaires avant de conclure, augmentant drastiquement la précision en mathématiques et programmation.
La capacité d'un LLM à déterminer quand une action externe est requise et à générer des arguments structurés pour invoquer des API, calculatrices ou bases de données avant de poursuivre.
Les scores de benchmarks, tarifs par token et limites évoluent rapidement. Vérifiez toujours les spécifications actuelles sur la documentation officielle du fournisseur.