Moonshot Kimi

Update 3 September 2026: As announced, Moonshot AI shut down kimi-k2.5 and the moonshot-v1 series on 31 August 2026; per the documentation, calls now return a 404 error. Anyone still using these IDs must migrate to kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed or kimi-k2.6 – the open K2.5 weights remain available for self-hosting. Moonshot has not released a new model since the K3 weight release.

Important Notice for Enterprises: innFactory AI Consulting from Rosenheim recommends Moonshot Kimi for the DACH region either self-hosted or through an inference provider with EU data centers. When using Moonshot’s own Kimi API, data is processed in China.

Kimi K3: open weights since July 27, 2026

Moonshot AI unveiled Kimi K3 on July 16, 2026, and published the full weights on Hugging Face (moonshotai/Kimi-K3) on July 27, 2026 – 96 safetensors shards totalling roughly 1.56 TB. That makes K3 one of the largest openly released models to date.

Architecture and capabilities (vendor-reported):

• 2.8 trillion parameters total, 104 billion active per token
• 896 experts, 16 activated per token, plus 2 shared experts
• 93 layers (69 KDA + 24 Gated MLA), 160,000-token vocabulary
• Context window: 1,048,576 tokens (1M)
• Native text, image and video input via the MoonViT-V2 vision encoder (401M parameters)
• Reasoning is always on: the model returns reasoning_content, with three effort levels (low, high, max; default max). In multi-turn use the complete assistant message including reasoning must be sent back (“preserved thinking history”)
• Tool calling through structured API responses
• Shipped as MXFP4 weights with MXFP8 activations, produced by quantization-aware training from the SFT stage onward – no separate quantization pass required
• Recommended inference engines: vLLM, SGLang, TokenSpeed

License: no longer Modified MIT

K3 is not released under MIT or Apache 2.0 but under a bespoke Kimi K3 License (tagged license:other on Hugging Face). Downloading, self-hosting, fine-tuning and quantizing are explicitly permitted, and commercial use is allowed by default. Two thresholds matter:

• Offering K3 to third parties as a service while exceeding USD 20M group revenue across any twelve consecutive months requires a separate agreement with Moonshot AI.
• Above 100M monthly active users or USD 20M monthly revenue, the license requires the words “Kimi K3” to be shown in the user interface.

Purely internal use that is never exposed to third parties is exempt from these obligations. For typical enterprise self-hosting little changes – but anyone reselling Kimi K3 as part of their own SaaS product should have the license text reviewed before rollout. The K2.x line stays under the Modified MIT License.

Official API pricing

According to the vendor the price is flat across the full one-million-token window, with no long-context surcharge. That puts K3 well above the K2.x line. Third-party providers price independently; a faster Kimi K3 Fast variant is offered by some gateways at a premium over the standard model.

Benchmarks: vendor figures and independent measurements

For K3 Moonshot reports, among others, GPQA Diamond 93.5, Terminal-Bench 2.1 88.3, BrowseComp 91.2, DeepSWE 67.5, SciCode 58.7, AA-LCR 74.7, DeepSearchQA (F1) 95.0, MMMU-Pro 81.6 and MathVision 94.3.

Independently, Artificial Analysis measures an Intelligence Index of 57, placing K3 at rank 3 – behind Fable 5 and GPT-5.6 Sol, on a comparable level with Opus 4.8 and GPT-5.5. Further figures from the same run: GDPval-AA v2 Elo 1668, AA-Briefcase Elo 1547, AutomationBench-AA 53%. On efficiency: 132M output tokens across the benchmark suite, about 21% fewer than K2.6 (166M), at roughly USD 0.94 cost per task.

Where the sources diverge: Artificial Analysis lists K3 at rank 3 or 4 depending on the evaluation date, while Arena.ai places it at the top of its coding ranking. The wider gap is on factuality: on the AA-Omniscience benchmark the hallucination rate rose from 39% (K2.6) to 51% (K3), while accuracy also improved (33% to 46%). The model answers correctly more often, but recognises less often when it does not know. That metric does not appear in Moonshot’s own charts. For knowledge-intensive applications without source grounding this is the key caveat – we recommend using K3 in such scenarios only with retrieval and source verification.

Hardware for self-hosting

The 1.56 TB of weights are the real hurdle. Under vLLM roughly 1.68 TB of VRAM are required. Realistic configurations start at one node with 8x B300 (2,304 GB) or four nodes with 8x H100 each (2,560 GB). Because the weights already ship as MXFP4, the usual “quantize FP16 down” route is no longer available – that headroom has already been spent. Running K3 on consumer hardware, a single GPU or via CPU offload is not practical.

If you want to run Kimi on your own infrastructure without maintaining such a cluster, Kimi K2.7 Code (~1T parameters, 256K context) is considerably cheaper.

Kimi K2.7 Code (June 2026)

On June 12, 2026, Moonshot AI released Kimi K2.7 Code – a coding-focused successor to Kimi K2.6. The model is available as open weights under the Modified MIT License and focuses on software development and agentic coding pipelines.

• ~1T parameter MoE architecture, ~32B active, 384 experts (8 active + 1 shared), 61 layers, 256K context
• MoonViT vision encoder (400M), forced thinking mode, OpenAI-compatible API
• Around 30% fewer thinking tokens than K2.6 (per Moonshot)
• Also offered on the Kimi platform as kimi-k2.7-code-highspeed (~180 tokens/s)

Benchmark caveat: The scores cited come from Moonshot’s own evaluation suites. Independent third-party verification is not available for K2.7 Code to the same extent as for K3, so these numbers should be interpreted with appropriate caution.

Kimi K2.6 (April 2026)

On April 20, 2026, Moonshot AI made Kimi K2.6 generally available. Compared with K2.5 the generation mainly improved agentic coding and long-horizon workflows (vendor-reported):

• SWE-Bench Pro: 58.6% (K2.5: 50.7%)
• Terminal-Bench 2.0: 66.7% (K2.5: 50.8%)
• BrowseComp (Agent Swarm): 86.3% (K2.5: 78.4%)
• Toolathlon: 50.0% (K2.5: 27.8%)
• Agent Swarm: up to 300 parallel sub-agents and 4,000 coordinated steps
• 1T parameter MoE, 32B active, 256K context, native INT4 quantization
• Open weights on Hugging Face (moonshotai/Kimi-K2.6)

K2.6 remains usable but is superseded by K2.7 Code for coding scenarios and by K3 as an all-round model.

Key Strengths

Very long context

• Kimi K3: 1,048,576 tokens – roughly 2,500 text pages, with no price surcharge for long inputs
• K2.x line: 256,000 tokens – roughly 640 text pages
• Suitable for extensive document analysis and multi-step reasoning across long conversations

Mixture-of-Experts architecture

• K3: 2.8T parameters total, 104B active, 896 experts (16 active + 2 shared), 93 layers made up of 69 KDA and 24 Gated MLA blocks
• K2.x: 1T parameters total, 32B active, 384 experts (8 active), Multi-Head Latent Attention
• Stable training via the MuonClip optimizer (K2 line)

Native multimodality

• K3: MoonViT-V2 vision encoder (401M parameters), text, image and video input in a single model
• K2.5/K2.6/K2.7 Code: MoonViT (400M parameters), processes images, videos, PDFs and text natively
• K3 vendor figures: MMMU-Pro 81.6, MathVision 94.3
• K2.5 vendor figures: MMMU-Pro 78.5%, MathVision 84.2%, OCRBench 92.3%, OmniDocBench 88.8%

Coding and agentic tasks

• K3 (vendor-reported): Terminal-Bench 2.1 88.3, DeepSWE 67.5, SciCode 58.7, BrowseComp 91.2, DeepSearchQA (F1) 95.0
• K2.6 (vendor-reported): SWE-Bench Pro 58.6%, Terminal-Bench 2.0 66.7%, Toolathlon 50.0%
• K2.5 (vendor-reported): LiveCodeBench v6 85.0% pass@1, MMLU-Pro 87.1%
• Native tool calling, multi-step task decomposition (“think → act → re-think”)
• Agent Swarm: up to 300 parallel sub-agents (K2.6), up to 100 (K2.5)

Important Considerations

GDPR assessment

Moonshot AI is headquartered in Beijing. Anyone using the official Kimi API or kimi.com transfers personal data to the People’s Republic of China. There is no adequacy decision by the European Commission for China. Such a transfer is therefore only permissible via the instruments in Articles 44 et seq. GDPR – in practice usually standard contractual clauses under Art. 46(2)(c) GDPR, supplemented by a transfer impact assessment evaluating the legal situation in the recipient country and government access to data. In addition, a data processing agreement under Art. 28 GDPR, a review of subprocessors, retention periods and technical safeguards are required.

Note: This is a factual summary of the legal framework and does not constitute legal advice. Whether a specific deployment is permissible depends on your role, the categories of data, contractual arrangements and safeguards in the individual case. For legal assessments we work with specialised law firms.

In practice this means: for personal or confidential business data the simpler route runs through the open weights – either self-hosted in the EU or via an inference provider with EU data centers. Prompts then never leave the EU and the question of a transfer to China does not arise in the first place.

Routes for EU deployment

• Self-hosting: Full data control, no third-country transfer. Very hardware-intensive for K3 (see above), considerably more accessible for K2.7 Code.
• EU inference providers: Nebius Token Factory has carried Kimi K3 since launch day and operates data centers in Finland and France with an EU data residency option (SOC 2 Type II, ISO 27001). DPA, subprocessors and retention must be reviewed case by case.
• Microsoft Foundry: Kimi K3 has been bookable through Foundry since July 31, 2026, but is served by Fireworks AI rather than as a first-party Azure endpoint. The actual data flow and region depend on the chosen deployment.
• No native offering on Amazon Bedrock or Google Vertex AI. AWS instead documents a self-managed deployment via SageMaker HyperPod or Amazon EKS.
• Some gateways offer K3 with a zero-data-retention option and routing to US providers – that resolves the China question but not the question of transfers to the United States.

Technical Highlights

Model Variants

• Kimi K3: Flagship (July 2026), 2.8T parameters, 1M context, multimodal, Kimi K3 License
• Kimi K2.7 Code: Coding specialist (June 2026), ~1T parameters, 256K context, Modified MIT – the more realistic self-hosting option
• Kimi K2.6: Long-horizon coding, agent swarm, 256K context
• Kimi K2.5: Multimodal, 256K context – API shut down on 31 August 2026, weights still on Hugging Face
• Kimi-K2-Thinking / -Instruct / -Base: API variants discontinued on May 25, 2026, weights still on Hugging Face

Licensing

• Kimi K3 License for K3 – open weights with revenue-triggered obligations for redistribution as a service
• Modified MIT License for the K2.x line
• Full weights on Hugging Face, reference code and benchmarks on GitHub (MoonshotAI/Kimi-K3)

Integration with CompanyGPT

If you want to run Kimi in your own infrastructure or with an EU provider, we support you with integration into CompanyGPT – our GDPR-compliant chat platform for enterprises. It lets you plug in open models such as Kimi alongside other providers without handing prompts to third countries.

Availability

Official API (data processed in China):

• platform.kimi.ai with OpenAI-compatible format
• Current models: kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed, kimi-k2.6
• Kimi K3: USD 3.00 / 1M input, USD 0.30 cached input, USD 15.00 / 1M output
• End-user access additionally via kimi.com

Open weights:

• Hugging Face: moonshotai/Kimi-K3 (~1.56 TB), moonshotai/Kimi-K2.7-Code, moonshotai/Kimi-K2.6, moonshotai/Kimi-K2.5
• GitHub: MoonshotAI/Kimi-K3
• Deployment via vLLM, SGLang and TokenSpeed

Third-party inference (selection):

• OpenRouter, Fireworks AI (day 0), Together AI, Baseten, Nebius Token Factory (day 0), Modal, DigitalOcean, Parasail, Databricks
• Microsoft Foundry since July 31, 2026 via Fireworks AI
• Pricing, regions and retention policies differ per provider and must be checked individually

Deprecated / discontinued:

• kimi-k2 series: discontinued on May 25, 2026
• kimi-latest: discontinued on January 28, 2026
• kimi-thinking-preview: discontinued on November 11, 2025
• kimi-k2.5 and the moonshot-v1 series: shut down on 31 August 2026, calls return 404 per Moonshot’s documentation

Our Recommendation (as of 3 September 2026)

With the weight release on July 27, 2026, the situation for Kimi K3 changed fundamentally: the model can now also be operated inside the EU – in your own data center or with an EU inference provider. That opens a GDPR path which did not exist before.

For most enterprises in the DACH region: Kimi K2.7 Code, self-hosted. At around 1T parameters and 256K context it runs on ordinary GPU server hardware, ships under the permissive Modified MIT License and covers coding and agent workflows very well.

For demanding all-round tasks: Kimi K3 – but realistically only via an EU inference provider unless you have a cluster with ~1.7 TB of VRAM available. For knowledge-intensive applications, the elevated hallucination rate (AA-Omniscience: 51% vs. 39% for K2.6) should be mitigated with retrieval, source citations and verification steps.

Not recommended for personal or confidential data: the direct API via platform.kimi.ai or kimi.com, because processing takes place in China and the requirements of Art. 44 et seq. GDPR would have to be met separately.

Alternative: Companies without their own GPU infrastructure and without an interest in open weights will find EU-hosted models such as Anthropic Claude or Google Gemini in the respective EU regions simpler to adopt.