Skip to content
FrontierDot

Kimi model timeline: K1.5 to K3.1 Pro, licenses and context (2025–2026)

Updated

Live timeline · last updated · 2 entries in the update log

Moonshot’s Kimi line runs from a January 2025 research model through Kimi K3 in July 2026 to Kimi K3.1 Pro in October 2026. This page lists each dated release we could tie to a Moonshot page, a GitHub README, a Hugging Face card or an official X post, and it separates a promise to release weights from a date we can actually point at.

As of October 25, 2026 the newest Kimi model with a dated announcement is Kimi K3.1 Pro, dated October 24, 2026 on the research blog index. The platform models page lists kimi-k3.1-pro and kimi-k3.1-pro-highspeed first, and still lists kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6. Full K3.1 Pro weights are promised by November 4, 2026 and were not verified as public on the check day. K2.7 Code still has no release date on that page, so it is not a node.

Read on 2026-10-25: the Kimi research blog index, the K3.1 Pro post, the official @Kimi_Moonshot posts from October 23, 2026 UTC, the K2, K2 Thinking, K2.5, K2.6 and K3 posts, the K1.5 and K2 and K3 GitHub READMEs, Hugging Face cards for the K2 family and K3, the Kimi platform models page, and the platform homepage pricing for K3.1 Pro. Scores and the “first,” “most capable” and “best” wording are Moonshot’s. We did not run the models.

Timeline

  1. The Kimi research blog index lists “Kimi K1.5: Scaling Reinforcement Learning with LLMs” on this day and points at the MoonshotAI/Kimi-k1.5 repository. The README introduces Kimi k1.5 as an o1-level multi-modal model and says the team scaled the context window of reinforcement learning to 128k. The technical report is arXiv 2501.12599. The README we read does not say the weights were released, so this node is a research release, not an open-weight release. The old kimi.com post slug for K1.5 now redirects to the blog index, which is why the index is the date source.

    Kimi research blog index

  2. Moonshot introduces Kimi K2, a mixture-of-experts model with 32 billion activated parameters and 1 trillion total parameters. The post says the team is open-sourcing Kimi-K2-Base and Kimi-K2-Instruct, and it calls K2 a reflex-grade model without long thinking. The GitHub README repeats the 32 billion and 1 trillion figures and says a 1 trillion parameter mixture-of-experts model was pre-trained on 15.5 trillion tokens. The Hugging Face card for moonshotai/Kimi-K2-Instruct names the license modified-mit. The license file is titled Modified MIT License and is copyright 2025 Moonshot AI.

    Kimi: Kimi K2

  3. The research blog index titles this day’s post Kimi-K2-Instruct-0905. The Hugging Face README for moonshotai/Kimi-K2-Instruct-0905 calls it the latest, most capable version of Kimi K2, keeps 32 billion activated parameters and 1 trillion total parameters, and says the context window has been increased from 128k to 256k tokens. The card uses the same modified-mit license name as the July K2 card. We date the node from the blog index, not from the model name alone, though the name and the index agree.

    Hugging Face: Kimi-K2-Instruct-0905 README

  4. Moonshot posts Kimi K2 Thinking and calls it the best open-source thinking model. The post says the model can execute up to 200 to 300 sequential tool calls, and that it is live in chat mode on kimi.com and on the API. The Hugging Face card uses the modified-mit license name. We do not cite a platform changelog line for this day: the English changelog we read does not contain a November 6 release sentence, so the blog post is the source.

    Kimi: Kimi K2 Thinking

  5. The Kimi platform models page lists kimi-thinking-preview as discontinued on this day. This is an availability change, not a new model. The page is a current document read on October 21, 2026, and it records the retirement date in its own retirement table. We do not describe what kimi-thinking-preview contained, because this page does not.

    Kimi platform: models

  6. Moonshot introduces Kimi K2.5 and calls it the most powerful open-source model to date. The post says it continued pretraining over approximately 15 trillion mixed visual and text tokens, and that the model is natively multimodal. It describes an agent swarm of up to 100 sub-agents, up to 1,500 tool calls, and up to 4.5 times faster than a single agent. Those speed and swarm figures are Moonshot’s. The model is available via Kimi.com, the Kimi App, the API and Kimi Code, with modes Instant, Thinking, Agent and Agent Swarm (Beta). The Hugging Face card uses modified-mit. A paper at arXiv 2602.02276 is not the release date.

    Kimi: Kimi K2.5

  7. The models page lists kimi-latest as discontinued on this day, the day after the K2.5 post. As with the November retirement, the page states the date and the model id, and it does not restate K2.5’s capabilities. We keep retirement nodes because an id that still appears in old sample code may no longer answer.

    Kimi platform: models

  8. Moonshot says it is open-sourcing its latest model, Kimi K2.6, available via Kimi.com, the Kimi App, the API and Kimi Code. The Hugging Face card uses the modified-mit license name. This node does not state a context length. The models page, read later, describes kimi-k2.6 as a 256k vision and text model, but that sentence is not in the April 20 post, so it is recorded in the status section below rather than here.

    Kimi: Kimi K2.6

  9. The models page gives this exact day for the retirement of the kimi-k2 series: kimi-k2-0905-preview, kimi-k2-0711-preview, kimi-k2-turbo-preview, kimi-k2-thinking and kimi-k2-thinking-turbo. The English platform changelog only says May 2026 for the same retirement, so the day comes from the models page, not from the changelog. Open weights that were already downloaded are a separate question from whether the preview API ids still respond.

    Kimi platform: models

  10. Moonshot introduces Kimi K3, a 2.8 trillion parameter model with Kimi Delta Attention and Attention Residuals, native vision and a 1 million token context. The post says it activates 16 out of 896 experts, uses Stable LatentMoE, and has about 2.5 times the scaling efficiency of Kimi K2. Moonshot calls it the world’s first open 3T-class model. It is available on Kimi.com, Kimi Work, Kimi Code and the Kimi API. At launch the default is maximum thinking effort, with low- and high-effort modes to come later. API prices on the post are $0.30 per million tokens for cache-hit input, $3.00 per million for cache-miss input and $15.00 per million for output. Moonshot says the official API cache hit rate is above 90 percent in coding workloads. The post says full model weights will be released by July 27, 2026. That sentence is a promise, not a record that the weights shipped on July 27.

    Kimi: Kimi K3

  11. The models page lists kimi-k2.5 and moonshot-v1 as discontinued on this day. The July 2026 platform changelog, which has a month but not a day, says K3 is on the Open Platform API with a 1 million token context and that k2.5 and moonshot-v1 are no longer offered to newly registered users. The August 2026 changelog says those two ids return 404. The day in this node is the models page’s date. The September 2026 changelog is about billing and search, not a model release, so it has no node.

    Kimi platform: models

  12. The Kimi research blog index dates “Kimi K3.1 Pro” to this day. The post introduces Kimi K3.1 Pro as the first model in the K3.1 series and as Moonshot’s most capable 3T-class open-weight model to date, with a 1 million token context. It describes K3.1 Pro as a post-trained version of Kimi K3 with substantially larger post-training compute, and it also releases Kimi K3.1 Pro Highspeed at two to three times the speed of K3.1 Pro. Both are available the same day on Kimi.com and Kimi.ai, Kimi Work, Kimi Code and the Kimi API. The post says full model weights will be released by November 4, 2026. That sentence is a promise, not a record that the weights shipped. Moonshot’s own numbers on the post include Terminal-Bench 4.0 at 65.7 percent; those scores are Moonshot’s, not an independent audit. On October 23, 2026 at 17:15 UTC, the official @Kimi_Moonshot account posted that K3.1 Pro has 2.8 trillion total parameters and 104 billion active parameters in a mixture-of-experts design, a 1 million token context and multimodal support, with open weights by November 4, 2026. A follow-up post at 17:20 UTC the same day says “Meet Kimi K3.1 Pro.” We keep the node date from the research blog index and treat the October 23 UTC posts as the same announcement, not a second release day. The blog body says 3T-class and post-trained K3; the 2.8 trillion and 104 billion figures come from the official X post.

    Kimi: Kimi K3.1 Pro

How we date a Kimi release

Many Kimi posts do not print a date in the article body. The date we use for a research post is the date paired with that title on the Kimi research blog index. Each node still links the page that states what changed: the blog post, the model card or the platform models page. If those two disagree, the node says so.

A repository timestamp is not a release date. GitHub’s API did not return a created-at time for MoonshotAI/Kimi-K3 or MoonshotAI/Kimi-k1.5 when we asked, and a Hugging Face lastModified time only says when the card was last edited. Neither becomes a node. Model names that contain a number, such as 0905, are checked against the index rather than treated as proof by themselves.

The last check for this page is October 25, 2026. Prices, context lengths and which ids still answer can move without a new blog post. The models page is the place we used for retirements, because it prints the day. The English changelog often prints only the month.

What actually changed from K1.5 to K3.1 Pro

K1.5 is the odd one out. It is a research report about scaling reinforcement learning to a 128k context, with a paper. We found no sentence that releases weights. From K2 onward the posts say they are open-sourcing a model, and the cards name a license.

K2 on July 11, 2025 is the first open-weight node: 32 billion activated parameters, 1 trillion total, base and instruct checkpoints, and no long thinking. The September 5 card keeps those parameter counts and lengthens the context from 128k to 256k. K2 Thinking on November 6 adds a thinking model that Moonshot says can run 200 to 300 tool calls in sequence. K2.5 on January 27, 2026 is the native multimodal step, with continued pretraining on about 15 trillion mixed visual and text tokens, distinct from the 15.5 trillion text tokens named in the K2 README.

K2.6 on April 20 is an open-source release on the same four surfaces, Kimi.com, the app, the API and Kimi Code. We do not attach 256k to that day. K3 on July 16 is the scale change: 2.8 trillion parameters, 16 of 896 experts active, a 1 million token context and native vision. The efficiency comparison with K2, about 2.5 times, is Moonshot’s claim. So is the cache-hit rate above 90 percent. The dollar prices on that post are the ones to quote for launch day.

K3.1 Pro on October 24, 2026 is a post-trained K3 with substantially larger post-training compute, still at a 1 million token context, plus a Highspeed variant Moonshot says is two to three times faster. The official X post states 2.8 trillion total and 104 billion active parameters. The blog calls it the most capable 3T-class open-weight model to date. Product surfaces are Kimi.com and Kimi.ai, Kimi Work, Kimi Code and the API. On the platform homepage read on October 25, 2026, kimi-k3.1-pro is listed at ¥13 input, ¥68 output, ¥1.30 cache hit and ¥13 cache write per million tokens; kimi-k3.1-pro-highspeed at ¥26, ¥136, ¥1.30 and ¥26. Those yuan figures are the page’s. We did not verify a matching English dollar API rate card for K3.1 Pro.

Weights promised, and weights we will not date

The July 16 K3 post says the full model weights will be released by July 27, 2026. The GitHub README, as read on October 21, 2026, says the full Kimi K3 weights are released under the Kimi K3 License, copyright 2026 Moonshot AI. The Hugging Face card moonshotai/Kimi-K3 lists the license name kimi-k3 and the pipeline image-text-to-text. Those two pages are enough to say the K3 weights are public as of that read. They are not enough to say the upload happened on July 27. There is no node for that day.

The October 24 K3.1 Pro post, and the October 23 UTC X posts, say full model weights will be released by November 4, 2026. As of the October 25, 2026 check, we did not verify a public weight upload for K3.1 Pro. That day is not a node.

The same gap applies to any specialist Kimi post we did not open. The index includes PerceptionBench, Agent Swarm, WorldVQA, Kimi Vendor Verifier, Kimi-Researcher, Kimi-Dev, Kimi-Audio, Kimina-Prover Preview, Kimi-VL, and the Muon, MoBA and Mooncake posts. They are real titles with dates on the index. This timeline is the general model line, and a title is not a substitute for the post body, so they are not nodes.

What the API lists now

Read on October 25, 2026, the platform models page lists kimi-k3.1-pro and kimi-k3.1-pro-highspeed first, then kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6. It describes kimi-k3.1-pro as the first model in the K3.1 series and the strongest flagship to date, at a 3 trillion-class parameter scale, with native vision and a 1 million token context. The highspeed id is described at about two to three times the speed of K3.1 Pro. kimi-k3 is still listed as the previous flagship, with 2.8 trillion parameters. kimi-k2.6 remains vision and text, thinking and non-thinking, at 256k. kimi-k2.7-code is listed at 256k. The high-speed K2.7 id is still described at about 180 tokens per second, and up to 260 on a short context. Those speeds are the page’s figures. Neither K2.7 id has a release date on the page, so neither is a node.

The same page carries an availability notice: from October 31, 2026 the K2 series stops for new registrants, while already registered users are unaffected, and Moonshot plans to retire all K2 series models by the end of December. That is a forward notice, not a completed discontinuation day like August 31, so it is not a timeline node. Completed retirements on the page remain the November 11, 2025, January 28, 2026, May 25, 2026 and August 31, 2026 nodes above. If you are updating an example that still calls kimi-k2-thinking or moonshot-v1, the page says those ids are gone. Downloaded K2 weights under the Modified MIT License are a different artifact from a retired preview id. For current yuan API rates, the platform homepage is the page we re-read; see the K3.1 Pro figures in the change section above.

Licenses, as the cards name them

  • K1.5: no weight license stated on the README we read.
  • K2, K2-Instruct-0905, K2 Thinking, K2.5 and K2.6: Hugging Face license name modified-mit. The K2 license file is a Modified MIT License, copyright 2025 Moonshot AI.
  • K3: Kimi K3 License, copyright 2026 Moonshot AI, on the GitHub LICENSE file. Hugging Face license name kimi-k3. This is not the Modified MIT text.
  • K3.1 Pro: as of October 25, 2026 the post promises weights by November 4, 2026. We have not verified a public card or license file for those weights yet.
  • We did not compare either license clause by clause with Apache 2.0. The name on the card is the fact. The grant is in the file.

What we left out

  • An activated-parameter figure for K3 beyond 16 of 896 experts on the July post. The 104 billion active figure in the October 23 X post is for K3.1 Pro, not a second K3 number from that July page.
  • Amazon Bedrock availability. An AWS what’s-new URL exists. The fetch did not return the announcement body, so there is no Bedrock node and no price or date from that page.
  • Kimi K2.7 Code’s release day. The models page lists it and does not date it.
  • Any score from a third-party leaderboard, including MarkTechPost-style roundups. Moonshot’s own “best,” “first” and Terminal-Bench figures are quoted as Moonshot’s.
  • Dollar API prices for K3.1 Pro. The English membership pricing page we read lists dollar membership plans, not a matching API token rate card. The yuan figures we quote come from the platform homepage.
  • A completed weight-upload day for K3.1 Pro. Only the November 4, 2026 promise is on the record as of October 25, 2026.
  • A October 31, 2026 timeline node for K2. The models page prints a forward notice that new registrants lose K2 access from that day, and a plan to retire all K2 by the end of December. That is not a past discontinuation day.
  • Specialist models whose posts we did not read, listed in the weights section above.

Related timelines

The same format covers Claude, OpenAI, Gemini, Grok and Meta Muse. Labs covered in this batch are Mistral, DeepSeek and Alibaba Qwen. The Front blog index lists every post.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

What is the latest Kimi model?

Kimi K3.1 Pro, dated October 24, 2026 on the research blog index. On October 25, 2026 the platform models page lists kimi-k3.1-pro and kimi-k3.1-pro-highspeed first, and still lists kimi-k3, kimi-k2.6 and kimi-k2.7-code. K2.7 Code has no date on that page.

Is Kimi K3.1 Pro open weight?

The October 24 post and the October 23 UTC X posts promise full weights by November 4, 2026. As of October 25, 2026 we have not verified a public weight upload. The model is available on Kimi.com, Kimi Work, Kimi Code and the API.

Is Kimi K3 open weight?

The July 16 post promised weights by July 27, 2026. The GitHub README, read on October 21, 2026, says the weights are released under the Kimi K3 License. We do not have a verified upload day.

What license is Kimi K2 under?

The Hugging Face cards say modified-mit. The K2 license file is titled Modified MIT License, copyright 2025 Moonshot AI.

When was Kimi K1.5 released?

The Kimi research blog index dates the K1.5 post to January 20, 2025. The README describes a research model with a 128k reinforcement-learning context and does not say the weights were released.

Update log

  • — Added Kimi K3.1 Pro (blog index October 24, 2026; official X posts October 23 UTC). Refreshed status, FAQ, API list and the November 4, 2026 weight promise. Checked through October 25, 2026.
  • — First published. Checked Moonshot’s Kimi posts, model cards and the platform models page through October 21, 2026.

Sources