Kimi K3 and K3.1: Specs, Context Window, Pricing, Benchmarks and Availability

Kimi K3 is an open-weight, natively multimodal Mixture-of-Experts model from Moonshot AI with 2.8 trillion total and 104 billion active parameters and a 1,048,576-token context window, announced in July 2026, and Kimi K3.1 Pro is its post-trained successor and the first model in the K3.1 series. This profile collects the published facts about both in one place, together with the K3.1 Pro Highspeed serving variant: API identifiers, architecture, context and output limits, prices, licences, thinking controls, provider-reported benchmarks, availability, data handling and lifecycle status. It is written for developers, architects, buyers and governance teams who need checked numbers rather than launch coverage. Every fact comes from Moonshot AI's model cards, blog posts and API documentation, listed under Standards and references. Where Moonshot AI publishes nothing on a point, the page says so.

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. It uses Moonshot AI's public documentation and Hugging Face model cards as of November 2026. Model details change, so confirm against Moonshot AI's current documentation.

Key takeaways

• Three API models. kimi-k3, kimi-k3.1-pro and kimi-k3.1-pro-highspeed, the last described as two to three times the speed of K3.1 Pro.
• Same size, different attention. K3 and K3.1 Pro both have 2.8T total and 104B active parameters and 93 layers; K3.1 Pro replaces K3's 24 Gated MLA layers with Gated DSA layers for long-context efficiency.
• 1M context, 131,072 default output. Both models list 1,048,576 tokens of context; the K3.1 Pro quickstart sets max_completion_tokens to 131,072 by default and up to 1,048,576.
• K3 is $3.00 input and $15.00 output per million tokens. Cache hits are $0.30 per the K3 launch blog, and the Batch API bills K3 and K3.1 Pro at 60% of standard price.
• Open weights under Moonshot licences. K3 uses the Kimi K3 License and K3.1 Pro the Kimi K3.1 Pro License; neither model has a retirement notice as of 15 November 2026.

What is Kimi K3?

Kimi K3 is the third generation of Moonshot AI's Kimi models, after the K2 series (K2, K2.5, K2.6 and K2.7-Code). Its model card describes it as the first open 3T-class model and states about 2.5 times better overall scaling efficiency than Kimi K2. It is built on Kimi Delta Attention (KDA) and Attention Residuals, uses MXFP4 weights and MXFP8 activations from quantisation-aware training, and pairs the language model with a 401M-parameter MoonViT-V2 vision encoder.

Kimi K3.1 Pro is, in Moonshot AI's words, "the first model in our new K3.1 series": a post-trained version of K3 "with substantially larger post-training compute", trained with large-scale joint reinforcement learning across hundreds of domains. Moonshot AI calls it its most capable 3T-class model to date. K3.1 Pro Highspeed was released alongside it as a faster serving option of the same model line. All three are reasoning models whose thinking cannot be switched off. For the full Kimi line-up, see Moonshot AI Kimi Models.

Key facts at a glance

As of 2026-11-15. Sources: Kimi K3 and K3.1 Pro Hugging Face cards, K3 and K3.1 Pro blogs, K3.1 Pro quickstart, Kimi API pricing page.

Release and version history

As of 2026-11-15. Moonshot AI's blog posts do not state exact announcement days, so the table gives the dates the official pages do state.

How much does Kimi K3 cost?

As of 2026-11-15. Per million tokens. The K3 blog uses a "$" sign without naming the currency code. The pricing page's tables did not render in the text checked, so K3.1 Pro figures could not be read there.

• Flat billing: the quickstart describes pay-as-you-go pricing with no tiering by context length; cache writes, cache hits, cache misses and output are each billed per token.
• Cache blocks: the cache is stored in blocks and any partial block is billed as a cache miss.
• Taxes and files: prices exclude applicable taxes; file content extraction and storage are temporarily free.
• Access: API use unlocks after a successful top-up of at least $1, and cumulative top-up sets the account tier and rate limits (concurrency, RPM, TPM, TPD).

Worked example at K3 standard rates with no cache hits: 100,000 input tokens and 10,000 output tokens (reasoning included) cost $0.30 for input plus $0.15 for output, $0.45 in total. Thinking is always on, so reasoning tokens add to output. See Model Pricing Explained.

Context window and output limits

The K3.1 Pro card attributes its Gated DSA layers (indexer with 32 heads and top-K 2048) to more efficient long-context modelling. Moonshot AI's own BrowseComp settings compact context at 200K tokens for K3.1 Pro and 300K for K3, which shows the provider does not run its agent evaluations at the full window. See Context Windows Explained.

Modalities and tools

• Vision input: images as base64 data URLs; uploaded video referenced as ms://<file-id> with purpose video. The quickstart states public image URLs are not supported and content must be an array.
• Thinking: always on, returned as reasoning_content. In multi-turn and tool-call conversations the full assistant message, including reasoning_content and tool_calls, must be passed back unchanged.
• Structured output and tool control: json_schema with strict true, partial mode, tool_choice "required" on the first turn and dynamic tool loading through a system message carrying tools.
• Fixed sampling (K3.1 Pro): temperature 1.0, top_p 0.95, n 1 and zero presence and frequency penalties; omit them from requests.
• Web search: the quickstart says the built-in web search is being updated and not recommended for production in the near term.
• Compatibility: OpenAI-compatible and Anthropic-compatible APIs; the quickstart base URL is https://api.moonshot.ai/v1.
• Self-hosting engines: vLLM, SGLang and TokenSpeed recommended; Transformers with trust_remote_code, Docker Model Runner and community quantisations for llama.cpp, Ollama and LM Studio. The K3 blog recommends supernode configurations of 64 or more accelerators.

The K3 blog warns that K3 "may act proactively on ambiguous intent" and advises explicit behavioural constraints in the system prompt or AGENTS.md. See the AI Coding Agents Security Guide.

Where is Kimi K3 available?

As of 2026-11-15. Sources: K3 and K3.1 Pro blogs and model cards.

Self-hosting the weights moves operation, patching and access control to the organisation. See the Self-Hosted LLM Security Guide.

Access restrictions

Neither model is described as restricted to invited or verified organisations. API access requires a top-up of at least $1, and rate limits rise with cumulative top-up. In Kimi Enterprise, K3.1 Pro is off by default at launch and must be enabled by an administrator. Weights are downloadable under the Kimi K3 License and Kimi K3.1 Pro License, whose terms are in each repository's LICENSE file; the cards do not summarise them. On cyber capability, the K3.1 Pro blog states that Moonshot AI removed cyber-related data from both training stages and aims not to release open weights that extend the absolute frontier of cyber capability.

Benchmarks (provider-reported, dated)

As of 2026-11-15. All figures are provider-reported on the Hugging Face model cards, at reasoning effort max and temperature 1.0. Independent leaderboard results are not included because none were checked against their own published pages for this profile. Only rows with the same benchmark name on both cards are paired.

• K3.1 Pro only: Terminal-Bench 4.0 65.7, SWE Marathon v1.1 51.7, OSWorld V2.1 79.6, Automation Bench-AA 73.3, ExploitGym 20.7.
• K3 only: Terminal-Bench 2.1 88.3, OSWorld-Verified 84.8, MCPMark-Verified 94.5, Video-MME (with subtitles) 90.0.
• Hallucination tests: Kimi Hallu Bench (internal, lower is better) reports 5.4 without tools and 8.3 on the hard set with tools for K3.1 Pro; the blog says these deliberately hard prompts do not reflect overall daily error rates.
• Caveats on the cards: several benchmarks are internal, some Kimi results come from specific harnesses (Claude Code or Kimi Code), and comparator figures were chosen by Moonshot AI and are not reproduced here.

See How to Read LLM Benchmarks.

Model card and system card

• Model cards: Hugging Face cards for K3 and K3.1 Pro give architecture, licence, deployment recipes, the thinking contract and benchmark tables. The K3 card links a technical report (k3_tech_report.pdf) on GitHub.
• Stated limitations (K3 blog): quality may become highly unstable if a harness does not pass back all historical thinking or a session switches to K3 midway; K3 may take unexpected actions on minor issues or ambiguous intent; and Moonshot AI acknowledges a user-experience gap against leading proprietary models.
• Cyber safety statement (K3.1 Pro blog): on the internal xdev-bench-hard, a version without the data removal measures scored 90.5% pass@3 against 81.0% for the released K3.1 Pro, and the blog says cyber capability remains behind frontier proprietary models public for at least three months.
• System card: not stated by Moonshot AI on the pages checked (2026-11-15).

See Model Cards and System Cards Explained.

Data handling

• Training use: the API data security page states that API inputs and outputs are "not used to train or improve Kimi's models"; the ZDR page states enterprise data is not used for training by default.
• Default retention: not stated by Moonshot AI on the pages checked (2026-11-15). Data is used to fulfil the current request and is not persistently stored for training purposes.
• Zero data retention: available on request to enterprise customers through sales; prompts and responses are deleted once the response is returned, except where law requires retention. Automated content safety reviews still apply.
• Outside ZDR: Batch API jobs, images and videos sent through the File Upload API, security logs and billing records, and data processed by third-party models, connectors or plugins.
• Transport and files: all API traffic uses HTTPS/TLS; uploaded files can be deleted through the file API or the console. Encryption at rest and data location are not stated on the pages checked.

Because ZDR requests cannot be traced afterwards, the ZDR page advises keeping your own request logs. Kimi API keys are non-human identities with spend and data exposure; see the LLMjacking Guide.

Deprecation status

As of 15 November 2026, the Kimi API pricing page lists no deprecation or retirement date for kimi-k3, kimi-k3.1-pro or kimi-k3.1-pro-highspeed. The only lifecycle notice on the page covers the K2 series, which is closed to new registrations from 31 October 2026 with retirement planned by the end of December 2026; existing users are not affected until then.

See the LLM Deprecation and Retirement Tracker.

How Kimi K3 differs from other models in its family

As of 2026-11-15. Sources: model cards, blogs, quickstart and pricing page. Models listed by release order, newest first.

The documented differences between K3 and K3.1 Pro are post-training scale, the Gated DSA attention change, benchmark results and the published price (K3 only). The documentation checked does not state what Highspeed changes beyond speed. For cross-provider context see Open-Weight Frontier Models Compared and Mixture of Experts Explained.

Practitioner checklist

• Pin kimi-k3, kimi-k3.1-pro or kimi-k3.1-pro-highspeed explicitly and record the adoption date, since no dated snapshots are published.
• Confirm K3.1 Pro and Highspeed prices in the Kimi console and note the date checked before committing budgets.
• Use a harness that returns the full assistant message, including reasoning_content and tool_calls, on every turn; do not switch models mid-session.
• Remove temperature, top_p, n and penalty parameters from K3.1 Pro requests, and set max_completion_tokens below the 131,072 default where long answers are not needed.
• Write explicit boundaries in the system prompt or AGENTS.md before giving K3 models agentic tasks, with human approval for destructive actions.
• Request ZDR if required and keep Batch API jobs and file uploads out of ZDR-sensitive workflows.
• Read the Kimi K3 or Kimi K3.1 Pro License before commercial self-hosting.
• Store Kimi API keys in a secrets manager, never in client code, and rotate them on staff changes.
• Move remaining K2 series workloads to a K3 model before the planned end of December 2026 retirement.

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 K3.1 Pro quickstart (accessed 2026-11-15)
• Moonshot AI: Kimi API model pricing (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)

Related NHI Mgmt Group resources: Moonshot AI Kimi Models, Chinese vs US Frontier Models Compared, Open-Weight Model Security Guide, Frontier AI Models Compared.