Kimi is the family of large language models from Moonshot AI, a Chinese AI company that publishes open-weight Kimi models on Hugging Face and serves them through the Kimi API Platform and the Kimi apps; its newest model, Kimi K3.1 Pro, is a post-trained version of Kimi K3 with a 1M token context window. This page lists every current Kimi model, its API identifier, context window, licence, published price and status, plus the version history, the K2 series retirement schedule and an index of model cards and technical reports. Every fact comes from Moonshot AI's own model cards, blog posts, API documentation and pricing page. Where Moonshot AI publishes no value, the page says "Not stated by Moonshot AI" rather than estimating. Facts are as of 15 November 2026.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.
Key takeaways
- Kimi K3.1 Pro is the newest model. API name kimi-k3.1-pro, a Mixture-of-Experts design with 2.8T total and 104B active parameters, 1,048,576 tokens of context and text plus image input, according to its Hugging Face model card.
- Both K3 models publish open weights. Kimi K3 ships under the Kimi K3 License and K3.1 Pro under the Kimi K3.1 Pro License, both Moonshot-specific licences; the older K2.5, K2.6 and K2.7-Code cards use a Modified MIT License.
- Kimi K3 has a published API price. The K3 launch blog lists $0.30 per million cache-hit input tokens, $3.00 cache-miss input and $15.00 output; the Batch API bills both K3 models at 60% of the standard price.
- The K2 series is being retired. From 31 October 2026 K2 series models are no longer available to newly registered users, and Moonshot AI plans to retire all K2 series models by the end of December 2026.
- Thinking is always on for the K3 models. Reasoning depth is set with reasoning_effort (low, high or max, default max), and the API is described as OpenAI and Anthropic compatible.
What is Moonshot AI?
Moonshot AI publishes its models under the Kimi brand. The current API platform is platform.kimi.ai; the K2 series model cards point to the earlier address platform.moonshot.ai. Both describe OpenAI-compatible and Anthropic-compatible endpoints. Consumer and work products carry the same name: Kimi.com (also kimi.ai), the Kimi mobile app, the Kimi Work desktop app, the Kimi Code terminal tool and Kimi Enterprise for organisations.
Model names follow a generation and point-release pattern: K2, K2.5, K2.6, K2.7-Code, K3 and K3.1. Suffixes mark a variant: Pro for the main K3.1 model, Highspeed for a faster serving variant and Code for a coding-focused model. API identifiers are lower-case versions of the names (kimi-k3, kimi-k3.1-pro, kimi-k3.1-pro-highspeed); the K2 cards do not name an official API identifier. See LLM Naming and Versioning Explained for how these conventions work across providers.
Current Kimi models at a glance
As of 15 November 2026. Sources: Kimi API pricing page, Kimi K3.1 Pro quickstart and Hugging Face model cards. Models are listed newest first.
| Model | API ID or version | Category | Context | Licence | Status |
|---|---|---|---|---|---|
| Kimi K3.1 Pro | kimi-k3.1-pro | Generalist, multimodal agentic | 1,048,576 | Kimi K3.1 Pro License | Current |
| Kimi K3.1 Pro Highspeed | kimi-k3.1-pro-highspeed | Faster serving variant of K3.1 Pro | Not stated by Moonshot AI on the pages checked (2026-11-15) | Not stated by Moonshot AI | Current |
| Kimi K3 | kimi-k3 | Generalist, multimodal agentic | 1,048,576 | Kimi K3 License | Current |
| Kimi K2 series (K2.7-Code, K2.6, K2.5, K2) | Not listed on the pricing page | Generalist and coding | 262,144 | Modified MIT License (K2.5, K2.6, K2.7-Code cards) | Closed to new users from 2026-10-31; retirement planned by end of December 2026 |
The K3.1 Pro blog states that Enterprise administrators must turn K3.1 Pro on for their organisation, as it is off by default at launch. The Highspeed variant is described as delivering two to three times the speed of K3.1 Pro.
Key facts for the main Kimi models
As of 15 November 2026. Identical fields for every model; "Not stated by Moonshot AI" means not stated by Moonshot AI on the pages checked (2026-11-15). API IDs, context, licence and status are in the table above.
| Model | Released | Max output | Knowledge cutoff | Input / output | Open weights | Parameters | Positioning (provider wording) |
|---|---|---|---|---|---|---|---|
| Kimi K3.1 Pro | Exact date not stated by Moonshot AI; weights due by 2026-11-04 (blog) | max_completion_tokens default 131,072, settable up to 1,048,576 (quickstart) | Not stated by Moonshot AI | Text, image / text | Yes (moonshotai/Kimi-K3.1-Pro) | 2.8T total, 104B active, plus 401M MoonViT-V2 vision encoder | "Our most capable 3T-class model to date"; first model in the K3.1 series |
| Kimi K3 | Announced July 2026 (blog; exact day not stated); weights due by 2026-07-27 | Not stated by Moonshot AI | Not stated by Moonshot AI | Text, image (card also mentions video) / text | Yes (moonshotai/Kimi-K3) | 2.8T total, 104B active, plus 401M MoonViT-V2 vision encoder | Native multimodal agentic model; described as the first open model at 2.8T parameters |
| Kimi K2.7-Code | Not stated by Moonshot AI | Not stated by Moonshot AI | Not stated by Moonshot AI | Text, image, video / text | Yes (moonshotai/Kimi-K2.7-Code) | 1T total, 32B active, plus 400M MoonViT encoder | Coding model built upon Kimi K2.6 |
| Kimi K2.6 | Not stated by Moonshot AI | Not stated by Moonshot AI | Not stated by Moonshot AI | Text, image, video / text | Yes (moonshotai/Kimi-K2.6) | 1T total, 32B active, plus 400M MoonViT encoder | Multimodal model in the K2 series |
| Kimi K2.5 | Card changelog entry 2026-01-29; paper dated 2 February 2026 | Not stated by Moonshot AI | Not stated by Moonshot AI | Text, image, video / text | Yes (moonshotai/Kimi-K2.5) | 1T total, 32B active (384 experts, 8 per token) | "Visual Agentic Intelligence" (paper title) |
The K3 and K3.1 Pro cards give the same architecture: 93 layers (one dense), 69 Kimi Delta Attention layers, 896 experts with 16 selected per token plus 2 shared experts, a vocabulary of about 160K and MXFP4 weights with MXFP8 activations from quantisation-aware training. K3 lists 24 Gated MLA layers and K3.1 Pro lists 24 Gated DSA layers. Only part of the network is active per token, which is why each card gives a total and an active parameter count. See Mixture of Experts Explained.
API features listed in Moonshot AI documentation
| Feature | Kimi K3.1 Pro | Kimi K3 | K2 series |
|---|---|---|---|
| OpenAI and Anthropic compatible API | Listed | Listed | Listed (platform.moonshot.ai) |
| Thinking mode | Always on | Always on, returns reasoning_content | Not stated by Moonshot AI on the pages checked |
| reasoning_effort (low, high, max) | Listed, default max | Listed, default max (blog: max only at launch) | Not stated by Moonshot AI on the pages checked |
| Automatic context caching | Listed (TTL 5m or 1h) | Listed (TTL 5m or 1h) | Listed (default TTL 5 minutes) |
| Batch API at 60% of standard price | Listed | Listed | Not stated by Moonshot AI |
| Strict JSON Schema output, tool choice, partial mode | Listed (quickstart) | Not stated by Moonshot AI on the pages checked | Not stated by Moonshot AI on the pages checked |
| Self-hosting engines named on the card | vLLM, SGLang, TokenSpeed, Docker Model Runner | vLLM, SGLang, TokenSpeed | vLLM, SGLang |
The K3.1 Pro quickstart states that sampling values are fixed (temperature 1.0, top_p 0.95, n 1, presence and frequency penalty 0) and should be omitted from requests, and that vision input accepts base64 data or uploaded file IDs, not public URLs. See Reasoning Models Explained.
Kimi model families and naming
Moonshot AI's model lines, as of 15 November 2026, from the Hugging Face organisation page and model cards.
- K3.1 line: Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed. The blog describes K3.1 Pro as post-trained from Kimi K3 with "substantially larger post-training compute".
- K3 line: Kimi K3, the base of the K3.1 series. The Hub shows adapters, fine-tunes and quantisations derived from it by third parties.
- K2 line: Kimi K2 (Base and Instruct repositories), K2.5, K2.6 and K2.7-Code, the last built upon K2.6. All share a 1T total, 32B active Mixture-of-Experts design and a 256K (262,144 token) context.
- Earlier and smaller open models on Hugging Face: Kimi-VL-A3B-Thinking and Kimi-VL-A3B-Thinking-2506 (vision language) and Moonlight-16B-A3B-Instruct. These are not listed on the Kimi API pricing page.
When was each Kimi model released?
Moonshot AI's blog posts and model cards checked for this page do not state exact launch days for most models. The table gives the dates that the official pages do state, as of 15 November 2026.
| Date | Event | API ID |
|---|---|---|
| 2026-01-29 | Kimi K2.5 card changelog entry; paper "Kimi K2.5: Visual Agentic Intelligence" dated 2 February 2026 (arXiv 2602.02276) | Not stated by Moonshot AI |
| July 2026 (day not stated) | Kimi K3 announced and available on Kimi.com, Kimi Work, Kimi Code and the Kimi API | kimi-k3 |
| By 2026-07-27 | Kimi K3 full weights release date stated in the launch blog | moonshotai/Kimi-K3 |
| Date not stated | Kimi K3.1 Pro and K3.1 Pro Highspeed announced | kimi-k3.1-pro, kimi-k3.1-pro-highspeed |
| By 2026-11-04 | Kimi K3.1 Pro full weights release date stated in the launch blog | moonshotai/Kimi-K3.1-Pro |
| 2026-10-31 | K2 series closed to newly registered users | K2 series |
| End of December 2026 (planned) | Retirement of all K2 series models | K2 series |
Release dates for K2, K2.6 and K2.7-Code are not stated on their model cards. The K3.1 Pro weights were published on Hugging Face by the date checked.
How much do Kimi models cost?
As of 15 November 2026. Per million tokens. Sources: Kimi K3 launch blog and the Kimi API pricing page (platform.kimi.ai/docs/pricing/chat).
| Model | Input | Cached input | Output | Batch | Cache write | Notes |
|---|---|---|---|---|---|---|
| Kimi K3 | $3.00 (cache miss) | $0.30 (cache hit) | $15.00 | 60% of standard price | Billed separately by TTL (5 minutes or 1 hour); amount not shown in the text checked | Source: K3 launch blog; batch and cache rules from pricing page |
| Kimi K3.1 Pro | Not shown in the text of the pricing page checked (2026-11-15); the K3.1 Pro blog links to the pricing page without listing prices | 60% of standard price | Billed separately by TTL (5 minutes or 1 hour) | Source: pricing page and quickstart | ||
| Kimi K3.1 Pro Highspeed | Not stated by Moonshot AI on the pages checked (2026-11-15) | Not stated by Moonshot AI | Not stated by Moonshot AI | Source: quickstart | ||
| K2 series | Not shown in the text of the pricing page checked (2026-11-15) | Not stated by Moonshot AI | Not stated by Moonshot AI | Existing users only | ||
Billing rules: the K3.1 Pro quickstart describes flat pay-as-you-go billing with no context-length tiers, with cache writes, cache hits, cache misses and output each billed per token. Partial blocks that do not fill a whole cache block are billed as misses. The pricing page states that context caching is applied automatically to repeated request prefixes with a default TTL of 5 minutes, that prices exclude applicable taxes, and that file content extraction and file storage are temporarily free, with extracted content billed as model input. API access requires a successful top-up of at least $1, and rate limits (concurrency, RPM, TPM and TPD) depend on an account tier set by cumulative top-up amount.
Currency note: the K3 blog quotes prices in US dollars. The pricing page's tables did not render in the text checked, so its currency could not be read; confirm the current figure in the Kimi console before committing budgets. For how input, output, cached and batch pricing work in general, see Model Pricing Explained. To cap spend on keys, see How to set spend limits on LLM API keys.
Where are Kimi models available?
- First-party API: the Kimi API Platform (platform.kimi.ai), with OpenAI-compatible and Anthropic-compatible endpoints. The K2 cards refer to platform.moonshot.ai.
- Apps: the K3 blog lists Kimi.com, the Kimi app on iOS, Android and HarmonyOS, Kimi Work desktop (version 3.1.0 or later, Windows and Apple silicon Macs) and Kimi Code in the terminal, where K3 is selected with the /model command.
- Enterprise: Kimi Enterprise for organisations, with separate personal and organisation accounts; administrators enable K3.1 Pro for their organisation.
- Open weights: K3.1 Pro, K3, K2.7-Code, K2.6, K2.5 and K2 repositories are on Hugging Face under moonshotai.
- Third-party hosting: the K3.1 Pro card lists Hugging Face Inference Providers (DeepInfra) and HuggingChat. Cloud marketplaces and regions are not stated by Moonshot AI on the pages checked (2026-11-15).
Running open weights yourself moves security responsibility to you. See the Open-Weight Model Security Guide and the Self-Hosted LLM Security Guide.
Restricted-access Kimi models
As of 2026-11-15. Restricted-access research programmes for Kimi models: not stated by Moonshot AI on the pages checked (2026-11-15).
| Model | Access stated by provider | Data retention condition |
|---|---|---|
| Kimi K3.1 Pro (Enterprise) | Off by default at launch for Enterprise organisations; administrators must enable it | ZDR available for API customers (K3.1 Pro blog) |
| K2 series | Existing users only from 2026-10-31 | Not stated by Moonshot AI |
| Kimi K3, K3.1 Pro (API) | Public API after a top-up of at least $1; open weights downloadable | ZDR on request for enterprise customers |
What is the context window of Kimi models?
Kimi K3.1 Pro and Kimi K3 have a context window of 1,048,576 tokens (1M) on both the pricing page and their model cards. The K2 series has 262,144 tokens (listed as 256K on the cards). The K3.1 Pro quickstart allows max_completion_tokens up to the full 1,048,576, with a default of 131,072. See Context Windows Explained and LLM Context Windows Compared Across Providers.
Is any Kimi model open source?
Every Kimi model listed on this page publishes weights on Hugging Face. Licences differ by generation. Kimi K3 uses the Kimi K3 License, covering the code repository and the weights, and Kimi K3.1 Pro uses the Kimi K3.1 Pro License; the cards link to the licence files without summarising the terms, so read them before commercial use. Kimi K2.5, K2.6 and K2.7-Code use a Modified MIT License. Open weights are not the same as open training data or code. See Open-Weight vs Closed Models Explained.
Model cards and system cards
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| Kimi K3.1 Pro | huggingface.co/moonshotai/Kimi-K3.1-Pro | Not stated by Moonshot AI | Not stated by Moonshot AI (tech blog at kimi.ai/blog/kimi-k3-1-pro) |
| Kimi K3 | huggingface.co/moonshotai/Kimi-K3 | Not stated by Moonshot AI | k3_tech_report.pdf in github.com/MoonshotAI/Kimi-K3 (linked from card) |
| Kimi K2.7-Code | huggingface.co/moonshotai/Kimi-K2.7-Code | Not stated by Moonshot AI | Not stated by Moonshot AI |
| Kimi K2.6 | huggingface.co/moonshotai/Kimi-K2.6 | Not stated by Moonshot AI | K2.5 paper, arXiv 2602.02276 (linked from card) |
| Kimi K2.5 | huggingface.co/moonshotai/Kimi-K2.5 | Not stated by Moonshot AI | arXiv 2602.02276 |
The K3 blog lists stated limitations: generation quality may become unstable if a harness does not pass back all historical thinking content or if a session switches to K3 mid-stream, and the model may take unexpected actions when instructions are ambiguous, so it recommends explicit constraints in the system prompt or an AGENTS.md file. See Model Cards and System Cards Explained.
Benchmarks
As of 15 November 2026. Provider-reported scores from the Hugging Face model cards, at reasoning effort max. Independent results are kept in a separate column.
| Benchmark | Kimi K3.1 Pro (provider-reported) | Kimi K3 (provider-reported) | Independent leaderboard results |
|---|---|---|---|
| GPQA Diamond | 94.2 | 93.5 | None cited on this page |
| HLE-Full, with tools | 70.0 | 56.0 | None cited on this page |
| BrowseComp | 92.1 | 91.2 | None cited on this page |
| MMMU-Pro, with Python | 90.0 | 83.4 | None cited on this page |
| CharXiv (RQ), with Python | 93.4 | 91.3 | None cited on this page |
| APEX-Agents | 69.4 | 41.0 | None cited on this page |
| DeepSWE | 75.5 (v1.1) | 67.5 | None cited on this page |
Benchmark versions differ between the two cards in places, so only rows with the same benchmark name are shown side by side. The cards report evaluation settings of temperature 1.0 and top-p 0.95. See How to Read LLM Benchmarks.
Deprecations and retirements
As of 15 November 2026, from the Kimi API pricing page.
| Model | API ID | Deprecated or notice | Retirement or shutdown | Stated replacement |
|---|---|---|---|---|
| K2 series (K2.7-Code, K2.6, K2.5, K2) | Not listed on the pricing page | 2026-10-31: no longer available to newly registered users; existing users not affected | Planned by end of December 2026 | Not stated by Moonshot AI |
Moonshot AI describes the retirement as planned, so watch the pricing page for an exact date. Inventory where each K2 identifier is used before the end of December 2026. See the AI Model Retirement Guide and the LLM Deprecation and Retirement Tracker.
Data handling and security
- Training use: Moonshot AI's API data processing page states that API inputs and outputs are not used to train or improve Kimi models. The ZDR page states that by default the Kimi Open Platform does not use enterprise customer data to train models.
- Default retention: not stated by Moonshot AI on the pages checked (2026-11-15). The data processing page says API data is used to fulfil the current request and is not persistently stored for training purposes.
- Zero data retention (ZDR): available on request to enterprise customers through the sales team. Under ZDR, request content is processed only while the response is generated and deleted once it is returned, except where law requires retention; automated content safety reviews still apply.
- Not covered by ZDR: Batch API jobs, images and videos uploaded through the File Upload API, operational data such as security logs and billing records, and data processed by third-party models, connectors or plugins.
- Encryption and files: API traffic is encrypted in transit with HTTPS/TLS. Uploaded files can be deleted through the file API or the console. Encryption at rest and storage region are not stated by Moonshot AI on the pages checked (2026-11-15).
- Key handling: generic controls apply to every provider's API: keep keys out of code, scope and rotate them, and set spend limits.
See the LLMjacking Guide and the Foundation Model Selection Guide. For agentic coding use through Kimi Code or any other harness, see the AI Coding Agents Security Guide.
How to choose a Kimi model: factual criteria
- Licence: the Kimi K3 License or Kimi K3.1 Pro License for the K3 generation, Modified MIT for K2.5, K2.6 and K2.7-Code.
- Lifecycle: new accounts cannot use the K2 series, and its retirement is planned by the end of December 2026.
- Speed: kimi-k3.1-pro-highspeed is described as two to three times the speed of kimi-k3.1-pro.
- Price per token: K3's published API price is $3.00 input, $0.30 cached and $15.00 output per million tokens; confirm K3.1 Pro pricing in the console.
- Self-hosting: 2.8T total parameters for K3 and K3.1 Pro versus 1T for the K2 series; the cards name vLLM, SGLang and TokenSpeed.
- Context: 1M on K3 and K3.1 Pro, 262,144 on the K2 series.
- Retention requirements: ZDR is an enterprise request and excludes Batch API jobs and files uploaded through the File Upload API.
Practitioner checklist
- Confirm the exact model ID (kimi-k3, kimi-k3.1-pro or kimi-k3.1-pro-highspeed) on the Kimi API Platform before you build against it.
- Check the price per million input and output tokens in the Kimi console, and note the date you checked.
- Move any K2 series workload to a K3 model before the planned end of December 2026 retirement.
- Pass back all historical thinking content in multi-turn sessions, as the K3 blog recommends, and avoid switching models mid-session.
- Set explicit constraints in the system prompt or AGENTS.md before giving the model agentic tasks.
- Read the Kimi K3 or K3.1 Pro licence file before commercial self-hosting.
- Request ZDR if you need it, and keep Batch API jobs and file uploads out of ZDR-sensitive workflows.
- Store Kimi API keys in a secrets manager and rotate them; never embed them in client code.
Standards and references
- Moonshot AI: Kimi K3.1 Pro model card (accessed 2026-11-15)
- Moonshot AI: Kimi K3 model card (accessed 2026-11-15)
- Moonshot AI: Introducing Kimi K3 (accessed 2026-11-15)
- Moonshot AI: Kimi K3.1 Pro blog (accessed 2026-11-15)
- Moonshot AI: Kimi API model pricing (accessed 2026-11-15)
- Moonshot AI: Kimi K3.1 Pro quickstart (accessed 2026-11-15)
- Moonshot AI: Zero data retention (accessed 2026-11-15)
- Moonshot AI: Kimi API data processing and security (accessed 2026-11-15)
- Moonshot AI: Kimi K2.6 model card, with the Kimi K2.5 and Kimi K2.7-Code cards (accessed 2026-11-15)
Related NHI Mgmt Group resources: Zhipu Z.ai GLM Models, Alibaba Qwen Models, LLM Security Guide, Open-Weight Model Security Guide.