Moonshot Kimi

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 centres. When using the international Moonshot API, data is stored on servers in Singapore per the privacy policy, i.e. outside the EU. As of October 2026.

Moonshot AI, headquartered in Beijing, builds the Kimi models as mixture-of-experts architectures with open weights. The currently relevant models are Kimi K3.1 Pro (API only, weights announced), Kimi K3 (2.8T parameters, open weights), Kimi K2.7 Code and Kimi K2.6 (~1T parameters each). All information comes from Moonshot AI, Hugging Face and the cloud providers’ documentation; benchmarks are vendor-reported unless stated otherwise.

Current models at a glance

Kimi K3.1 Pro

Kimi K3.1 Pro (kimi-k3.1-pro) is the first model of the K3.1 series and, per Moonshot AI, a post-trained version of Kimi K3 with substantially more post-training effort. It has a context window of 1,048,576 tokens, understands text, images and video natively and supports tool calling. Reasoning is always on and configurable via reasoning_effort in three levels (low, high, max; default max). In multi-turn conversations and tool calls the complete assistant message including reasoning must be passed back. There is also kimi-k3.1-pro-highspeed at 2-3x the speed with the same intelligence, per Moonshot.

Moonshot highlights improvements in coding (including game engines such as Unity, Unreal Engine and Godot), 3D modelling and financial analysis. Selected vendor figures:

• Terminal-Bench 4.0: 65.7% (Moonshot lists GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4%)
• Kimi Code Bench 3.0 Long-horizon (internal benchmark): 61.2% versus 55.95% for K3
• BenchCAD (with tools): 92.7%
• OfficeQA Pro v2: 64.4
• Hallucination rate in the internal stress test “Kimi Hallu Bench Hard”: 8.3% versus 23.4% for K3. Moonshot stresses that this test deliberately contains error-prone cases and does not represent everyday error rates.

On cybersecurity, Moonshot states that it removed cyber-related data from mid-training and RL environments; on its internal xdev-bench-hard the model reaches 81.0% versus 90.5% without these measures.

API access is unlocked after a top-up (minimum USD 1). Per Moonshot, zero data retention is available for API customers. Moonshot has explicitly announced that the full weights will be released by 4 November 2026. Until then the licence text, architecture details and hardware requirements are not settled. Availability at a cloud provider or EU inference provider is not documented so far.

Kimi K3

Moonshot AI unveiled Kimi K3 on 16 July 2026, and published the full weights on Hugging Face (moonshotai/Kimi-K3) on 27 July 2026 – 96 safetensors shards totalling roughly 1.56 TB. Moonshot now positions K3 as the previous generation to K3.1 Pro; the API remains available.

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 – roughly 2,500 pages of text, with no surcharge for long inputs
• Native text, image and video input via the MoonViT-V2 vision encoder (401M parameters)
• Reasoning always on with three effort levels (low, high, max; default max)
• Tool calling via structured API responses
• Shipped as MXFP4 weights with MXFP8 activations from quantization-aware training – no downstream quantisation run
• Recommended inference engines: vLLM, SGLang, TokenSpeed

Benchmarks: 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. In the AA-Omniscience benchmark the hallucination rate rose from 39% (K2.6) to 51% (K3), while the hit rate also rose (33% to 46%): the model answers correctly more often but less often recognises when it does not know. For knowledge-intensive applications without source grounding we therefore recommend retrieval and source checks.

Kimi K2.7 Code

Kimi K2.7 Code (12 June 2026) is a coding-focused successor to Kimi K2.6 with open weights under the Modified MIT License.

• ~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
• Per Moonshot around 30% fewer thinking tokens than K2.6
• Also available on the Moonshot platform as kimi-k2.7-code-highspeed (~180 tokens/s)

The benchmarks come from Moonshot’s own evaluation suites; independent third-party verification is not available to the same extent as for K3.

Kimi K2.6

Kimi K2.6 (20 April 2026) improved on K2.5 mainly in 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 quantisation
• Open weights on Hugging Face (moonshotai/Kimi-K2.6)

K2.6 is superseded by K2.7 Code for coding and by K3 / K3.1 Pro as an all-round model.

Pricing

Moonshot API list prices per 1 million tokens (USD, excluding taxes). Per Moonshot the prices apply flat across the whole context window; no long-context surcharge is listed.

Context caching is billed automatically (default TTL 5 minutes). Third-party providers and cloud platforms set their own, differing prices.

Licence

• Kimi K3.1 Pro: licence not yet published; Moonshot speaks of open weights and announces the weights by 4 November 2026.
• Kimi K3: the custom Kimi K3 License (tagged license:other on Hugging Face). Downloading, self-hosting, fine-tuning, quantising and commercial use are generally permitted. Anyone offering K3 to third parties as a service with group revenue above USD 20M over twelve consecutive months needs a separate agreement with Moonshot AI. From 100M monthly active users or USD 20M monthly revenue, a visible “Kimi K3” attribution in the interface is required. Purely internal use is exempt; if you pass K3 on within your own SaaS product, have the licence text reviewed.
• Kimi K2.x (K2.7 Code, K2.6): Modified MIT License.

Self-hosting and hardware

For Kimi K3 the 1.56 TB of weights are the real hurdle. Under vLLM around 1.68 TB of VRAM is 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 ship as MXFP4 already, further quantisation is not an option. Consumer hardware, a single GPU or CPU offload are not practical. We can only document K3.1 Pro’s requirements once the weights are released; as a post-trained K3 a similar order of magnitude is plausible but unconfirmed.

Kimi K2.7 Code (~1T parameters, 256K context) is the far cheaper self-hosting option. Recommended inference engines: vLLM, SGLang and TokenSpeed; reference code is on GitHub (MoonshotAI/Kimi-K3).

Availability in the EU and GDPR

GDPR assessment

Moonshot AI is headquartered in Beijing. Per its privacy policy, the international Kimi platform (platform.moonshot.ai) is operated by Moonshot AI Pte. Ltd. in Singapore, and data is stored on servers in Singapore. Anyone using the API therefore transfers personal data to a third country. There is no adequacy decision by the European Commission for Singapore. kimi.com and the Chinese platform have their own privacy terms, which must be checked separately before use. 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. A data processing agreement under Art. 28 GDPR, a review of subprocessors, retention periods and technical safeguards are also required. Moonshot offers a zero-data-retention option for API customers; it does not change the place of processing.

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.

For personal or confidential business data the simpler route runs through the open weights – self-hosted in the EU or via an inference provider with EU data centres.

Routes for EU deployment

• Self-hosting: Full data control, no third-country transfer. Very hardware-intensive for K3, considerably more accessible for K2.7 Code.
• EU inference providers: Nebius Token Factory has carried Kimi K3 since launch day and operates data centres in Finland and France with an EU data residency option (SOC 2 Type II, ISO 27001). Review DPA, subprocessors and retention case by case.
• AWS Bedrock: Since 18 September 2026, Bedrock carries Kimi K3 as a native model (moonshotai.kimi-k3). Per the AWS model card it is not available in-region, only via US-Geo cross-region (us.moonshotai.kimi-k3) or Global cross-region (global.moonshotai.kimi-k3) inference; the latter routes requests to any supported AWS Region worldwide. The request can be made from eu-central-1 (Frankfurt), but processing is not fixed to the EU – this is not an EU residency option. Its predecessor Kimi K2.5 (moonshotai.kimi-k2.5) is different: per the AWS model card it remains active and available as an In-Region model, including in eu-north-1 (Stockholm) (eu-west-2 London is in the United Kingdom); AWS gives 27 January 2027 as the earliest end of life. Kimi K3.1 Pro is not listed in AWS documentation.
• Microsoft Foundry: Kimi K3 has been bookable through Foundry since 31 July 2026, served by Fireworks AI rather than as a first-party Azure endpoint. Pay-per-token is offered as Global Standard or Data Zone Standard, with Fireworks on Foundry offering data zones in the US only; per Microsoft, Fireworks on Foundry is currently excluded from the EU Data Boundary. Fireworks’ pay-per-token offerings for K2.5, K2.6 and K2.7 Code have been retired. Microsoft Learn also lists Kimi K2.7 Code, K2.6 and K2.5 as preview models sold by Azure, per the region table as Global Standard only and therefore without an EU data zone.
• Google Agent Platform: Google retired its last Kimi model-as-a-service endpoint (Kimi K2 Thinking) on 21 October 2026; the alternative is self-deployment from Model Garden. K3 and K3.1 Pro are not listed. AWS also documents a self-managed deployment via SageMaker HyperPod or Amazon EKS.
• Other inference providers (selection): OpenRouter, Fireworks AI, Together AI, Baseten, Modal, DigitalOcean, Parasail, Databricks. Prices, regions and retention differ; some gateways offer zero data retention with routing to US providers – that resolves the Singapore question but not transfers to the United States.

Migration and breaking changes

• K3 to K3.1 Pro: Switch the model ID to kimi-k3.1-pro or kimi-k3.1-pro-highspeed. Reasoning stays always on, reasoning_effort (low/high/max) is a top-level field; in multi-turn use, return the complete assistant message including reasoning_content. Access requires a platform top-up.
• K2 series: From 31 October 2026 Moonshot no longer accepts new users for K2 models (K2.7 Code, K2.6) and plans to retire all K2 models by the end of December 2026. Existing integrations should migrate to K3.1 Pro (all-round); self-hosted coding workloads can stay on the open weights.
• K2.5 and moonshot-v1: Calls have returned a 404 error since 31 August 2026.

Model lifecycle

The weights of the older models (Kimi K2.5, Kimi-K2-Thinking, Kimi-K2-Instruct, Kimi-K2-Base) remain on Hugging Face for self-hosting; Kimi K2.5 also continues to run on AWS Bedrock, while Google retired the MaaS endpoint for K2 Thinking on 21 October 2026. Kimi K2.5 offered 256K context and native multimodality (LiveCodeBench v6 85.0%, MMLU-Pro 87.1%, agent swarm of up to 100 sub-agents per Moonshot); K2 Thinking added chain-of-thought reasoning.

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.

Our Recommendation

Kimi remains an option for the DACH region when the weights are operated yourself or by an EU provider.

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 well. Because Moonshot’s K2-series API is being retired by the end of 2026, this applies to the open weights, not to API access.

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

On AWS with a fixed EU region: On Bedrock, only Kimi K2.5 is available as an In-Region model in Stockholm – an option for existing AWS environments, albeit an older model generation.

Watch Kimi K3.1 Pro: Per Moonshot it is more capable and cheaper than K3, but today it can only be used via the Moonshot API (servers in Singapore). Whether and how it can be operated in the EU depends on the weights announced for 4 November 2026, the licence and the offerings of EU inference providers. Until then it suits only data without personal data and without confidentiality needs.

Not recommended for personal or confidential data: the direct Moonshot API and kimi.com, because processing takes place outside the EU (per the privacy policy, the API stores data in Singapore) and the requirements of Art. 44 et seq. GDPR would have to be met separately. Bedrock with Global or US cross-region is not an EU residency option either.

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.