Kimi Models: K3, K3.1 Pro and K2.7 Code Explained

Kimi is Moonshot AI's model family. Kimi K3 is a 2.8-trillion-parameter open-weight model with a 1,048,576-token context, text and image input and a custom licence. Its successor on the API, Kimi K3.1 Pro, is a post-trained version of K3 whose weights Moonshot says will be released by 4 November 2026. K2.7 Code is the older 1T coding model. Checked 31 October 2026.

The models

Facts below come from the Hugging Face cards, repository files and Moonshot's API documentation, read on 31 October 2026.

K3's architecture mixes Kimi Delta Attention, a linear attention variant, with gated multi-head latent attention, plus a design Moonshot calls Attention Residuals, and its Stable LatentMoE scales expert sparsity to 16 of 896. Moonshot claims about 2.5 times better scaling efficiency than Kimi K2. That is a vendor claim.

Which one to start with

The docs make the lineup clear. K3.1 Pro is "our most capable 3T-class open-weight model to date", with its weights "released by Nov 4". K3 is "previous-generation". The K3.1 Pro quickstart still contains unfinished placeholder text for one benchmark score ("[xx.x%]"), so we quote no K3.1 numbers. Check Moonshot's model card once it is final.

If you need weights today, K3 is the one you can download. If you can wait four days, K3.1 Pro's weights should be the better choice, since it is K3 with substantially more post-training. If you want a smaller coding model, K2.7 Code is the one, with the caveat in the next section about its API life.

Retirement dates you should know

Moonshot's model list gives a dated retirement schedule:

So K2.7 Code's hosted API has about two months left, and brand-new accounts cannot use it. Its weights stay on Hugging Face, so you can still self-host it indefinitely. For hosted use, plan a move to K3 or K3.1 Pro.

Licences

Kimi K3 uses the Kimi K3 License, which we read. In plain words:

• You may use, copy, modify, fine-tune, deploy and sell the software, including the weights.
• If you or an affiliate operate a Model-as-a-Service business (giving third parties inference or fine-tuning access with meaningful control over inputs, parameters or training data), and your group revenue exceeds 20 million US dollars over any 12 months, you need a separate agreement with Moonshot first.
• If a commercial product built on it has more than 100 million monthly active users or more than 20 million dollars in monthly revenue, "Kimi K3" must be shown on its interface.
• Internal use, and use through Moonshot's official products or certified inference partners, is outside the Model-as-a-Service rule.

Kimi K2.7 Code uses a Modified MIT License. Its file is a plain MIT text plus one display clause: products above 100 million monthly active users or 20 million dollars monthly revenue must show "Kimi K2.7 Code" on their interface. No Model-as-a-Service clause appears in that file.

For most builders both are workable. A hosted-inference reseller with real revenue is the one case that needs a conversation. Compare with the stricter or looser terms for Mistral, Qwen and GLM on our open source LLM page.

API access and prices

Moonshot's API platform serves the models at https://api.moonshot.ai/v1 in OpenAI format, and offers a Responses API and an Anthropic-compatible Messages API. Details from its docs:

• Access. Flagship models unlock after a successful top-up of at least $1. Cumulative top-ups set your account tier and rate limits. A cumulative $5 top-up earns a $5 voucher.
• Caching. Context caching bills automatically for repeated prefixes, with a default five-minute time to live. For K3 series models, cache writes are billed by TTL (5 minutes or 1 hour), and a cache hit is billed at the hit price with no extra write charge.
• Batch. kimi-k3.1-pro and kimi-k3 support batch inference at 60 percent of the standard price.
• Reasoning effort. The K3.1 Pro docs describe low, high and max.
• Context. All K3 series models support 1,048,576 tokens, and K2 series 262,144.
• Search and fetch. Separate web search and URL fetch endpoints are billed apart from tokens.

Per-token prices sit in a table on Moonshot's pricing page that is drawn by script and was empty when we fetched the page, so we quote no figures. The page does say billing is pay-as-you-go with no subscription, that cache writes are billed by TTL, and that the Batch API costs 60% of the standard price for K3.1 Pro and K3. Check the live page before budgeting.

Published scores

All scores are vendor-reported on the Kimi K3 and K2.7 Code cards. Moonshot's K3 numbers use max reasoning effort and temperature 1.0.

Kimi K3 (max effort)

Source: Kimi K3 model card. Moonshot ran K3 with the Kimi Code harness and for several rivals used the best score across harnesses, some taken from vendors' own posts or leaderboards, and notes that Fable 5 results may involve fallbacks. One more oddity: the card's ProgramBench row shows 77.8 for K3, while DeepSeek's and Z.ai's cards print 17.5 for the same model on a stricter "almost solved" reading. The tests differ, which is exactly why cross-card comparisons are shaky.

Kimi K2.7 Code

Source: Kimi K2.7 Code card, tested in Kimi Code CLI with thinking on, against GPT-5.5 in Codex and Opus 4.8 in Claude Code at xhigh effort. This card is candid: K2.7 Code trails both rivals on most rows. Its value was open weights and a small active size, not the top score.

Running the weights

There is no workstation route for K3. If you want to self-host something Kimi-sized without a cluster, you cannot. Use the API or choose a smaller model; see Qwen3.8-27B or gpt-oss-120b. For renting GPU nodes see Runpod, Modal and Lambda. Moonshot's licence mentions "certified inference partners" without naming them.

One operational rule from the K3 card: it was trained with preserved thinking history, so multi-turn and tool-calling clients must pass the complete assistant message back, including the reasoning content. A client that strips it will see weaker results.

Which coding agents can use Kimi

Moonshot's documentation has a setup guide for each of these.

If you use Claude Code, check your shell for old ANTHROPIC variables first. The guide warns that stale exports interfere with the new configuration. See also our Claude Code page.

How the line has moved

Hugging Face upload dates give a clean timeline of Moonshot's open releases: Kimi K2-Thinking in November 2025, K2.5 on 1 January 2026, K2.6 on 14 April 2026, K2.7 Code on 11 June 2026 and K3 on 13 June 2026. That is five open releases in about seven months, and the hosted side has retired older names just as fast: kimi-k2 on 25 May, K2.5 and the moonshot-v1 line on 31 August, and K2 series scheduled for the end of December. The lesson for buyers is the same as for DeepSeek: pin exact model IDs, expect them to disappear within a year, and budget a migration test whenever a new flagship lands.

K3 and K2.7 Code were published two days apart, which tells you they are different bets. K2.7 Code is the 1T model tuned for coding with a 256K window and native INT4, while K3 is the scale play. Moonshot's own K2.7 Code card places it behind GPT-5.5 and Opus 4.8 on most rows, so the company is not pitching it as a frontier model, only as a capable open one.

Highspeed variants

Moonshot sells faster versions of two models. kimi-k3.1-pro-highspeed is described as the same intelligence at about two to three times the speed of K3.1 Pro. kimi-k2.7-code-highspeed is the same model as K2.7 Code, with an output speed of about 180 tokens per second and up to 260 in short-context cases, per its quickstart. Those speeds are vendor claims. The price premium for highspeed was not visible in our fetch of the pricing page, which is one more reason to read it before you pick.

For an interactive coding agent, latency is a large part of the feel, so highspeed is probably the right default for a human at the keyboard. For unattended overnight runs, the standard model at a lower price makes more sense, or the batch API at 60 percent of the standard rate for work that can wait.

Setting up Claude Code with Kimi

Moonshot's guide asks you to write environment variables into the env block of ~/.claude/settings.json, then restart Claude Code. The shape of the file is:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.moonshot.ai/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "<your Kimi API key>",
    "ANTHROPIC_MODEL": "kimi-k3[1m]"
  }
}

That is a trimmed version. The guide's full example also sets the default Opus, Sonnet, Haiku and Fable tier models and the subagent model, so that every tier Claude Code picks resolves to a Kimi model. Two details matter. The [1m] suffix selects the million-token window, and the Haiku tier can point at the cheaper kimi-k2.7-code, which keeps small background calls off the flagship. After restarting, run a short task and confirm in the Moonshot console that the calls arrived.

Before any of that, remove stale ANTHROPIC exports from your shell files. Otherwise the shell values can win over the settings file and your calls go to the wrong place, which is the most common failure the guide lists.

What Kimi is good for, from the cards

Moonshot's own descriptions point at three strengths, and it is fair to treat them as claims to test rather than facts.

• Long engineering sessions. The K3 card says it sustains long sessions with minimal human oversight, navigates large repositories and orchestrates terminal tools, with examples as far afield as GPU kernels, compilers and chip design.
• Visual loops. Because K3 reads screenshots and video, it can use visual feedback in game development, front-end and CAD work.
• Knowledge work. The card describes deep research with charts and dashboards, and its table includes finance, legal and office benchmarks such as GDPval-AA v2, OfficeQA Pro and Harvey Lab-AA.

If your work is pure repository coding, the Terminal-Bench and DeepSWE rows are the ones to weigh. If it is document-heavy, the vision and office rows matter more. None of them replaces a trial on your own tasks.

Who should pick Kimi

Limits and gotchas

• Hardware floor. Terabyte-scale weights put K3 beyond almost every private deployment.
• Pricing visibility. The docs' price table was blank in our fetch, so confirm before use.
• Pending release. K3.1 Pro weights are promised, not yet published.
• Retirement. K2 series is closing to new users and retiring by the end of 2026.
• Top-up gate. Flagship API access requires a prepaid balance of at least $1.

Our take

K3 is the biggest openly downloadable model we have documented, and K3.1 Pro looks set to extend that if the weights arrive on schedule. Moonshot's documentation is the most agent-friendly of the Chinese labs, with named guides for Claude Code, Codex, OpenCode and OpenClaw. The licence is easy for products and awkward only for resellers. Until the weights and the price table are visible, treat K3.1 as a promise and K3 as the thing you can actually use. See best open source LLM for coding and GLM models for the closest competitor.

Product names are used descriptively. We are not affiliated with Moonshot AI.