HostYourAI
Product
EU RouterDecision models like JevHostYourAI CodeOpenAI-compatible APIAnthropic-compatible APIModel GardenDedicated InstancesPlaygroundFine-tuning (Loes)Earn with your GPU
Solutions
Use cases
HostYourAI CodeLLM InferenceRAG pipelinesChatbotsAI agentsFine-tuning
Industries
GovernmentHealthcareFinanceLegal
Models
Kimi K3DeepSeek V4 ProDeepSeek V4 FlashGLM 5.3Kimi K2.7 CodeQwen3.5 397BLlama 3.3 70BMistralMiniMax M3All models →
Compare
Azure OpenAIAWS BedrockClaude APIChatGPTOpenAICompare costs
Resources
DocumentationSavings checkGuide: migrate to the EU RouterGuide: deploy your own LLM (vLLM)Guide: build RAG on EU GPUsModel catalogNewsSecurityFAQ
Pricing News
NL EN DE
Login Book Demo Get Started
Model garden Router · on request Dedicated · on request

Kimi K2 Thinking

This model runs as a dedicated deployment on large GPUs and isn't in the shared playground by default. Get in touch and we'll set it up for you.

Kimi K2 Thinking is the latest, most capable version of open-source thinking model. Starting with Kimi K2, we built it as a thinking agent that reasons step-by-step while dynamically invoking tools. It sets a new state-of-the-art on Humanity's Last Exam (HLE), BrowseComp, and oth...

Request access ← All models
moonshotai/Kimi-K2-Thinking On request
text->text · moonshotai · sovereign EU
Runs on EU infrastructure operated by European companies; US marketplace capacity is never part of this chain. Full chain
1026B
Parameters
262K
Context window
4899GB
Minimum VRAM
POST /api/v1/chat/completions On request

Specifications

Parameters 1026B
Context window 262,144 tokens
Minimum VRAM 4899 GB
Architecture DeepseekV3ForCausalLM (vLLM)
License other
Modality text->text
Released November 2025
Publisher moonshotai ↗

Pricing

Shared router · per token
On request
Not available on the shared router. Pricing on request as a dedicated GPU deployment.
Dedicated GPU · per hour
On request
Dedicated deployment, from 4899 GB of VRAM. Billed per GPU-hour.

Shared EU router, pay-per-token, scale-to-zero. Dedicated GPU deployments are billed hourly, see pricing.

Bench index

Bench is our quality index for every model, 0 to 100. It rests on public measurements (Epoch AI with about 40 benchmarks, LiveBench, LMArena and more) and is refreshed every morning. Where nobody measures, it shows an estimate with a range.

70.0 / 100 · Measured

Measured by Epoch AI across 9 benchmarks, with a range.

Range 66.7 to 71.7 · Rank 82 of 277 at Epoch AI

Sources: Epoch AI ↗ · LiveBench ↗ · LMArena ↗ · Refreshed daily, last on 16-10-2026

LiveBench · 61.6
Reasoning 63.5
Coding 67.4
Agentic coding 38.3
Mathematics 81.1
Data analysis 52.3
Language 66.5
Instruction following 62.0
Every benchmark (9)
OTIS Mock AIME 2024-2025 83.0
GPQA diamond 79.0
METR Time Horizons 59.2
WeirdML 42.8
FrontierMath-2025-02-28-Private 37.6
Terminal Bench 35.7
CL-bench 17.6
Chess Puzzles 15.8
FrontierMath-Tier-4-2025-07-01-Private 0.0
LiveBench · All tasks
Reasoning
theory of mind 57.7
zebra puzzle 46.3
spatial 92.0
logic with navigation 58.0
Coding
code generation 71.8
code completion 63.0
Agentic coding
javascript 40.0
typescript 20.0
python 55.0
Mathematics
AMPS Hard 96.0
integrals with game 52.0
math comp 96.1
olympiad 80.3
Data analysis
consecutive events 15.7
tablejoin 45.1
tablereformat 96.1
Language
connections 88.0
plot unscrambling 43.4
typos 68.0
Instruction following
paraphrase 63.4
simplify 50.9
story generation 67.4
summarize 66.5

Call it now

Drop-in replacement for OpenAI: change only the base URL and API key. The Anthropic format (/v1/messages) is supported too.

curl https://hostyourai.com/api/v1/chat/completions \
  -H "Authorization: Bearer hyai-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/Kimi-K2-Thinking",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Frequently asked questions

Can I run Kimi K2 Thinking in the EU?

Yes. HostYourAI runs Kimi K2 Thinking on GPUs in European datacenters via vLLM. Prompts and outputs never leave the EU and there is no US cloud provider in the chain.

Is hosting Kimi K2 Thinking GDPR-compliant?

Yes. All processing happens inside the EU, a Data Processing Agreement (DPA) is available and the subprocessor list is public. Open-source weights also mean: no training on your data.

How much does Kimi K2 Thinking cost?

Kimi K2 Thinking needs several GPUs at once, so it runs as a dedicated deployment billed per GPU-hour rather than per token. Tell us your volume and we will work it out with you.

Is the API OpenAI-compatible?

Yes. You use the standard OpenAI SDKs with a custom base URL (https://hostyourai.com/api/v1). The Anthropic Messages API is supported as a drop-in as well.

More models from Moonshotai

Kimi K3

Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. It is a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.

1M context View model →
Kimi K2.7 Code

Kimi K2.7 Code is a coding-focused agentic model built upon Kimi K2.6. With substantial improvements on real-world long-horizon coding tasks, it strengthens end-to-end task completion across complex software engineering workflows while improving token efficiency, reducing thinking-token usage by approximately 30% compared with Kimi K2.6.

1027B 262K context View model →
Kimi K2.6

Kimi K2.6 is an open-source, native multimodal agentic model that advances practical capabilities in long-horizon coding, coding-driven design, proactive autonomous execution, and swarm-based task orchestration.

1027B 262K context View model →
Kimi K2.5

Kimi K2.5 is an open-source, native multimodal agentic model built through continual pretraining on approximately 15 trillion mixed visual and text tokens atop Kimi-K2-Base. It seamlessly integrates vision and language understanding with advanced agentic capabilities, instant and thinking modes, as well as conversational and agentic paradigms.

1027B 262K context View model →
Kimi Linear 48B A3B Base

Kimi Linear is a hybrid linear attention architecture that outperforms traditional full attention methods across various contexts, including short, long, and reinforcement learning (RL) scaling regimes. At its core is Kimi Delta Attention (KDA)—a refined version of Gated DeltaNet that introduces a more efficient gating mechanism to optimize the use of finite-state RNN memory.

49B View model →
Kimi Linear 48B A3B Instruct

Kimi Linear is a hybrid linear attention architecture that outperforms traditional full attention methods across various contexts, including short, long, and reinforcement learning (RL) scaling regimes. At its core is Kimi Delta Attention (KDA)—a refined version of Gated DeltaNet that introduces a more efficient gating mechanism to optimize the use of finite-state RNN memory.

49B View model →

Request access

Kimi K2 Thinking isn't available by default yet. Leave your details and we'll arrange a dedicated deployment.

Request access