Important Notice for Enterprises: innFactory AI Consulting from Rosenheim recommends Moonshot Kimi exclusively for self-hosted deployments. When using the API, data is processed in China - only self-hosting in your own infrastructure ensures GDPR compliance for the DACH region (Germany, Austria, Switzerland).
Latest Update: Kimi K3 (July 2026)
On July 16, 2026, Moonshot AI unveiled Kimi K3 – at 2.8 trillion parameters the company’s largest model to date and more than twice the size of the previous flagship K2.6. The MoE architecture activates 16 of 896 experts per token, the context window grows to 1 million tokens, and image input is supported natively.
- Artificial Analysis leaderboard: debut at #3 – behind Claude Fable 5 and GPT-5.6 Sol, ahead of Claude Opus 4.8 and GPT-5.5 High in many individual benchmarks
- Leader on the Arena.ai frontend code arena
- According to Moonshot’s own long-horizon evaluation, Elo 1547 (+732 vs. K2.6) – a vendor figure, to be interpreted with appropriate caution
- ~21% fewer output tokens than K2.6, currently only one reasoning level (“max”)
- API pricing: ~$3.00 / 1M input, ~$15.00 / 1M output – well above K2.x prices
Important for EU deployments: The open weights are not yet available at release (announced by July 27, 2026). Until then, K3 runs exclusively via website and API with data processing in China – meaning there is currently no GDPR-compliant path for K3. For productive EU use, self-hosted Kimi K2.6 remains our recommendation until the K3 weights are published and the hardware requirements for self-hosting are clarified.
Kimi K2.7 Code (June 2026)
On June 12, 2026, Moonshot AI released Kimi K2.7 Code - a coding-focused successor to the flagship Kimi K2.6 (released April 20, 2026). As usual, the model is available as open weights under the Modified MIT License, with a clear focus on software development and agentic coding pipelines.
- 1T parameter MoE architecture, ~32B active, 384 experts, 256K context window
- Forced thinking mode and OpenAI-compatible API
- Kimi Code Bench v2: 21.8% - an improvement over K2.6
- Pricing: ~$0.95 / 1M input tokens, ~$4.00 / 1M output tokens
Important benchmark caveat: The scores cited above come from Moonshot’s own evaluation suites. No independent third-party verification was available at the release date, so these numbers should be interpreted with appropriate caution.
For EU deployments, self-hosting on your own EU infrastructure via the open weights on Hugging Face remains the cleanest GDPR path. Managed clouds currently offer related models (AWS Bedrock: Kimi K2 Thinking / K2.5, EU via Frankfurt; Azure Foundry: Kimi K2.6 / K2.5 in Preview), but K2.7 Code itself is best run self-hosted for now. As the general flagship, Kimi K2.6 remains our recommendation; K2.7 Code primarily addresses coding use cases.
Kimi K2.6 (April 2026)
On April 20, 2026, Moonshot AI released Kimi K2.6 as a generally available model - eight days after the K2.6 Code Preview. The new generation significantly outperforms K2.5 on agentic coding and long-horizon workflows:
- SWE-Bench Pro: 58.6% (vs. 50.7% for K2.5) - beats GPT-5.4 (57.7%) and Claude Opus 4.6 (53.4%)
- Terminal-Bench 2.0: 66.7% (up from 50.8% for K2.5)
- BrowseComp (Agent Swarm): 86.3% (up from 78.4% for K2.5)
- Toolathlon: 50.0% (up from 27.8% for K2.5)
- Agent Swarm scaling: up to 300 parallel sub-agents and 4,000 coordinated steps
- Day-0 availability on Cloudflare Workers AI (
@cf/moonshotai/kimi-k2.6) - 1T parameter MoE architecture, 32B active, 256K context, native INT4 quantization
- Open-weight on Hugging Face (
moonshotai/Kimi-K2.6)
K2.6 is therefore currently our recommended Kimi model for self-hosting and agentic coding pipelines. The successor Kimi K3 was released on July 16, 2026 (see above) – as long as its open weights have not been published, K2.6 remains the first choice for EU operation.
Key Strengths
Ultra-Long Context (256K Tokens)
Kimi K2.5 offers one of the longest context windows in the open-source space:
- 256,000 tokens - approximately 640 text pages
- Ideal for extensive document analysis
- Enables complex multi-step reasoning over long conversations
- Surpasses most available open-source LLMs
Mixture-of-Experts Architecture
Innovative MoE architecture for maximum efficiency:
- 1 trillion parameters total
- Only 32 billion parameters active per request
- 384 experts, 8 activated per token
- Multi-Head Latent Attention (MLA) for efficient processing
- Stable training through MuonClip optimizer
- MMLU-Pro: 87.1% - Competitive with leading proprietary models in general knowledge
Native Multimodality (K2.5)
Kimi K2.5 is fully multimodal-trained with outstanding benchmark results:
- MoonViT Vision Encoder (400M parameters)
- Processes images, videos, PDFs, and text natively
- ~15 trillion mixed visual/text tokens in training
- MMMU-Pro: 78.5% - State-of-the-art visual multimodal understanding
- MathVision: 84.2% - Mathematics with visual integration
- OCRBench: 92.3% - Best OCR performance of all tested models
- OmniDocBench: 88.8% - Excellent document understanding
- Code generation from UI designs possible
State-of-the-Art Coding
Outstanding performance in programming tasks (2026 results):
- LiveCodeBench v6: 85.0% pass@1 - highest score among open-source models
- SWE-Bench Verified: 76.8% success rate (agentic coding) - close to proprietary frontier models
- Full-stack capabilities: Frontend, Backend, DevOps
- Debugging and automatic code optimization
- Precise function/tool calling for API integration
- AIME 2025: 96.1% (Math Olympiad Level) - excellent mathematical problem-solving
Agentic Intelligence
Specialized in autonomous reasoning workflows:
- Native tool-calling support
- Multi-step task decomposition (“think → act → re-think”)
- Agent Swarm System: Up to 100 parallel sub-agents for complex tasks
- Nearly 100% precision in API-based tool calls
- Kimi K2 Thinking: Transparent chain-of-thought processes
- GPQA-Diamond: 87.6% - Graduate-level reasoning performance
Important Considerations
Privacy Considerations
Important for Enterprises: Moonshot AI is a Chinese company. When using the API, data is processed on servers in China. For sensitive enterprise data, we strongly recommend self-hosting.
- Chinese Company: Server location China (with API usage)
- GDPR: No explicit GDPR compliance with API usage
- No EU Regions: Not available through AWS Bedrock, Azure AI, or Vertex AI
- Recommendation: Self-hosting in your own EU infrastructure
Self-Hosting as Solution
The open-source model can be operated in your own infrastructure:
- All data remains under your control
- No dependency on external APIs
- Full GDPR compliance possible
- Hardware Requirements: 8x H100 GPUs for full 256K context (INT4)
- Support for vLLM, SGLang, TensorRT-LLM deployment
Resource Requirements
Kimi is a resource-intensive model:
- Full Context (256K): 8x NVIDIA H100 (INT4 quantization)
- Standard Context (128K): 4-6x high-end GPUs
- INT4, INT8, and FP16 quantization supported
- For smaller deployments: Use shorter context windows
Technical Highlights
Model Variants
Moonshot AI offers various specializations:
- Kimi K3: New flagship (July 2026), 2.8T parameters, 1M context – open weights announced (by July 27, 2026)
- Kimi-K2.6: Long-horizon coding, 300-agent swarm, 256K context – current self-hosting recommendation
- Kimi-K2.5: Multimodal, 256K context, predecessor of K2.6
- Kimi-K2-Thinking: Reasoning-focused, transparent thought processes
- Kimi-K2-Instruct: General-purpose chat, 128K context
- Kimi-K2-Base: Foundation model for custom fine-tuning
Open-Source Licensing
Transparent open-source availability:
- Modified MIT License - permissive usage
- Complete model weights on Hugging Face
- Active community on GitHub
- Technical reports and benchmarks publicly available
Integration with CompanyGPT
If you wish to operate Kimi in your own infrastructure, we support you with integration into CompanyGPT - our GDPR-compliant chat platform for enterprises.
Availability
API Access:
- Public API available via platform.moonshot.ai
- Kimi K3 available via website and API since July 16, 2026 (~$3.00 / 1M input, ~$15.00 / 1M output)
- OpenAI-compatible API format
- Server location: China
- Not recommended for sensitive EU data
Self-Hosting:
- Model Weights: Hugging Face (moonshotai/Kimi-K2.6, moonshotai/Kimi-K2.5)
- GitHub Repository: MoonshotAI/Kimi-K2
- Deployment via vLLM, SGLang, TensorRT-LLM
- New: Kimi K2.6 on Cloudflare Workers AI with Day-0 support (since April 20, 2026)
- Kimi K2.5 on Cloudflare Workers AI (since March 19, 2026)
- NVIDIA NIM integration available
- Recommended for GDPR-compliant usage
Our Recommendation
Kimi K3 marks a leap for open-weight models: #3 on the Artificial Analysis leaderboard, 1M context, and native multimodality. However, as long as the open weights are not published (announced by July 27, 2026), the only access is via the China-based API – making it unusable for DACH enterprises for now. Kimi K2.6 remains technically impressive with state-of-the-art performance in agentic coding (SWE-Bench Pro 58.6%), long-horizon workflows, and multimodal tasks.
For DACH enterprises, we recommend exclusively self-hosting in your own EU infrastructure. When using the API via platform.moonshot.ai, data is processed in China, which is GDPR-critical.
Best Choice: Self-hosted Kimi-K2.6 for agentic coding pipelines and multimodal applications or Kimi-K2-Thinking for complex reasoning tasks - each in your own infrastructure with full data control. Once the K3 weights are released, we will assess its self-hosting suitability (at 2.8T parameters, hardware requirements are likely to be substantial).
Alternative: For companies without their own GPU infrastructure, we recommend EU-hosted alternatives like Anthropic Claude or Google Gemini via corresponding EU regions.