Kimi K3, K3.1 Pro, and the K2 line (November 2026): release dates, API pricing, open weights, size, and where to use them
Quick answer
Moonshot AI's newest Kimi model is Kimi K3.1 Pro, dated October 24, 2026 on Moonshot's blog: a 2.8-trillion-parameter open-weight model with a 1,048,576-token context, at $2 in and $10 out per million tokens on the Kimi API. Kimi K3, from July 16, 2026, is now the previous flagship at $3 / $15, and the K2 line (K2, K2 Thinking, K2.5, K2.6, K2.7 Code) is retired or closing: Moonshot plans to retire every K2 model by the end of December 2026. Every figure here is from Moonshot's own pages, read on November 18, 2026.
Free plan: no credit card, 1 published page, 50 form entries a month.
On this page
- Kimi K3 and K3.1 Pro: the current models
- Kimi K2, K2.5, K2.6, K2.7 Code, K3, and K3.1 Pro: release dates and status
- Kimi K2.5 vs K2.6 vs K3: what changed
- Kimi K3 size: parameters and download
- Kimi K3 pricing: the API
- Kimi K3 subscription: the Kimi plans
- Kimi K3 benchmarks: vs Fable 5, Opus 5.5, and GPT-6
- Is there a Kimi K3.1 beyond K3.1 Pro, or a K2.x still coming?
- Build and publish a page with Kimi
- Frequently asked questions
- Sources
You searched for Kimi K3 and found K3, K3.1 Pro, K2.5, K2.6, and K2.7 Code all mentioned at once. This page sorts them: which Kimi model is current, what each release changed, what it costs, and which ones you can still call.
Kimi is the model family of Moonshot AI. Every fact below comes from Moonshot's own pages: its research blog, the Kimi API platform's model list, pricing page, and changelog, and its Hugging Face model cards and license files, read on November 18, 2026 (Japan time).
Kimi K3 and K3.1 Pro: the current models
Kimi's K3 series, read on November 18, 2026. API prices are US dollars per million tokens.
| Kimi K3.1 Pro | Kimi K3 | |
|---|---|---|
| Status | Current flagship: "Kimi's most capable flagship to date" | "Kimi's previous-generation flagship" |
| Dated on Moonshot's blog | October 24, 2026 | July 16, 2026 |
| Weights on Hugging Face | Yes. The blog promised them "by Nov 4, 2026"; the repository dates from November 2 | Yes. The blog promised them "by July 27, 2026" |
| License | Kimi K3.1 Pro License | Kimi K3 License |
| Parameters | 2.8T total, 104B activated | 2.8T total, 104B activated |
| Download size | 1,561 GB of weight files | 1,561 GB of weight files |
| Context | 1,048,576 tokens | 1,048,576 tokens |
| Input | Text and images | Text, images, and video |
| API name | kimi-k3.1-pro, and kimi-k3.1-pro-highspeed | kimi-k3 |
| API price, input / output | $2 / $10. Highspeed: $4 / $20 | $3 / $15 |
| Cached input | $0.20 | $0.30 |
Kimi K3.1 Pro is "Post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains", and "the first model in the K3.1 series." It costs a third less than K3 on the API. The Highspeed variant is "same intelligence, about 2-3× the speed" at twice the price. Batch jobs on both K3-series models are "billed at 60% of the standard price."
Kimi K2, K2.5, K2.6, K2.7 Code, K3, and K3.1 Pro: release dates and status
Every Kimi K-series release, read on November 18, 2026. The date is the one on Moonshot's research blog, or the Hugging Face repository's creation date where the blog has no post.
| Model | Date | Parameters | Context | License | On the Kimi API today |
|---|---|---|---|---|---|
| Kimi K2 (K2-Instruct) | July 11, 2025 | 1T total, 32B activated | 128K | Modified MIT | Retired in May 2026 |
| Kimi K2-Instruct-0905 | September 5, 2025 | 1T total, 32B activated | 256K | Modified MIT | Retired in May 2026 |
| Kimi K2 Thinking | November 6, 2025 | 1T total, 32B activated | 256K | Modified MIT | Retired in May 2026 |
| Kimi K2.5 | January 27, 2026 | 1T total, 32B activated | 256K | Modified MIT | Retired in August 2026: calls "return a 404" |
| Kimi K2.6 | April 20, 2026 | 1T total, 32B activated | 256K | Modified MIT | Existing users only; "Deprecation expected in late December 2026" |
| Kimi K2.7 Code | Repository created June 11, 2026 | 1T total, 32B activated | 256K | Modified MIT | Existing users only; "Deprecation expected in late December 2026" |
| Kimi K3 | July 16, 2026 | 2.8T total, 104B activated | 1,048,576 | Kimi K3 License | Yes, kimi-k3 |
| Kimi K3.1 Pro | October 24, 2026 | 2.8T total, 104B activated | 1,048,576 | Kimi K3.1 Pro License | Yes, kimi-k3.1-pro |
Moonshot's pricing page carries the notice for the whole K2 line: "starting October 31, 2026, K2 series models will no longer be available to newly registered users; existing users are not affected. We plan to retire all K2 series models by the end of December 2026". Its changelog tells K2 and K2.5 users: "Please migrate to Kimi K3."
The weights don't retire. Every model in the table is still on Hugging Face under its license, for running on your own servers.
Kimi K2.5 vs K2.6 vs K3: what changed
- Kimi K2 → K2.5 (January 2026). K2.5 added vision: "a native multimodal agentic model built through continual pretraining on approximately 15 trillion mixed visual and text tokens atop Kimi-K2-Base", with instant and thinking modes. Context went from K2's 128K to 256K.
- K2.5 → K2.6 (April 2026). Same architecture: "Kimi-K2.6 has the same architecture as Kimi-K2.5". Moonshot describes K2.6 as advancing "long-horizon coding, coding-driven design, proactive autonomous execution, and swarm-based task orchestration."
- K2.6 → K2.7 Code (June 2026). "a coding-focused agentic model built upon Kimi K2.6", "reducing thinking-token usage by approximately 30%".
- K2 line → K3 (July 2026). A new, larger model: 2.8T parameters against 1T, a 1,048,576-token context against 262,144, and video input. K3 "is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes)" and "activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2." Moonshot calls it "the world's first open 3T-class model".
- K3 → K3.1 Pro (October 2026). Same size, trained further with reinforcement learning, and a lower API price.
The license changed with K3 too. K2-line weights use a Modified MIT License whose one added clause is a display rule: a product with "more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue" must show the model's name. The Kimi K3 and K3.1 Pro licenses keep that rule and add a second: a "Model as a Service" business whose revenue "exceeds 20 million US dollars" over 12 months "must enter into a separate agreement with Moonshot AI" before commercial use.
Kimi K3 size: parameters and download
Model size, read on November 18, 2026. Download size is the total of the weight files in each Hugging Face repository.
| Model | Total parameters | Activated per token | Weight files |
|---|---|---|---|
| Kimi K3.1 Pro | 2.8T | 104B | 1,561 GB |
| Kimi K3 | 2.8T | 104B | 1,561 GB |
| Kimi K2.7 Code, K2.6, K2.5 | 1T | 32B | 595 GB each |
| Kimi K2-Instruct-0905 | 1T | 32B | 1,029 GB |
K3 ships quantized: Moonshot trains it with "MXFP4 weights with MXFP8 activations for broad hardware compatibility", and K2.5 to K2.7 Code ship with native INT4 quantization. To serve K3 or K3.1 Pro yourself, Moonshot's cards name three inference engines: vLLM, SGLang, and TokenSpeed.
Kimi K3 pricing: the API
Kimi API prices per million tokens, read on November 18, 2026.
| Model | Input | Output | Cached input | Context |
|---|---|---|---|---|
kimi-k3.1-pro | $2.00 | $10.00 | $0.20 | 1,048,576 |
kimi-k3.1-pro-highspeed | $4.00 | $20.00 | $0.20 | 1,048,576 |
kimi-k3 | $3.00 | $15.00 | $0.30 | 1,048,576 |
kimi-k2.7-code (existing users) | $0.95 | $4.00 | $0.19 | 262,144 |
kimi-k2.7-code-highspeed (existing users) | $1.90 | $8.00 | $0.38 | 262,144 |
kimi-k2.6 (existing users) | $0.95 | $4.00 | $0.16 | 262,144 |
- Pay as you go. The API "is billed on a pay-as-you-go basis, with no subscription plan."
- Unlocking K3. "Kimi K3 is a flagship model: it is unlocked after a successful top-up (minimum $1)."
- Cache writes. On K3-series models, a 5-minute cache write costs the input price and a 1-hour write twice that.
- Thinking. "K3 always has thinking mode enabled", with
reasoning_effortset tolow,high, ormax. - API formats. Moonshot's cards: "we provide OpenAI/Anthropic-compatible API".
Kimi K3 subscription: the Kimi plans
The Kimi app, Kimi Work, and Kimi Code run on a Kimi membership, billed apart from the API.
| Plan | Monthly | Yearly, per month | What it adds |
|---|---|---|---|
| Free | $0 | 1 agent task at a time | |
| Plus | $19 | $15 ($180 a year) | More agent credits, Kimi Code, Deep Research, website deployment |
| Pro | $39 | $31 ($372 a year) | 2x agent credits, "Access to K3.1 Pro Highspeed", Goal mode |
| Max | $99 | $79 ($948 a year) | 5x agent credits, "Up to 1M tokens of context for ultra-long conversations" |
| Ultra | $199 | $159 ($1,908 a year) | 10x agent credits |
Kimi's plans page carries a purchase notice: "K3.1 Pro requests may queue or fail at launch. Wait times vary with overall load." Per Moonshot's launch posts, both K3 and K3.1 Pro are "available today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API", and the Kimi app runs on iOS, Android, and HarmonyOS.
Kimi K3 benchmarks: vs Fable 5, Opus 5.5, and GPT-6
Moonshot published a table with each release. Each compares with the US models that were current that month, and every score is Moonshot's own report.
Kimi K3's card (July 2026), all at max effort.
| Benchmark, as Moonshot reports it | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 |
| Terminal-Bench 2.1 | 88.3 | 88.0 | 88.8 | 84.6 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 |
| SWE-Marathon | 42.0 | 35.0 | 39.0 | 40.0 |
| Kimi Code Bench 2.0 (Moonshot's own) | 72.9 | 76.9 | 64.8 | 71.7 |
Kimi K3.1 Pro's card (November 2026), all at max effort.
| Benchmark, as Moonshot reports it | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 65.7 | 66.4 | 58.2 | 57.9 |
| DeepSWE v1.1 | 75.5 | 74.2 | 74.1 | 67.4 |
| SWE Marathon v1.1 | 51.7 | 46.3 | 42.5 | 45.6 |
| HLE-Full with tools | 70.0 | 67.7 | 57.2 | 65.6 |
| Kimi Code Bench 3.0 Long-horizon (Moonshot's own) | 61.2 | 58.5 | 66.5 | 58.8 |
Moonshot's notes say K3.1 Pro "is evaluated with the Claude Code harness" and that other models' scores are "the maximum values reported in model blog posts". The Terminal-Bench team's own leaderboard, on the newer Terminal-Bench 4.1, lists Kimi K3.1 Pro (max) in Claude Code at 59.7% ± 4.2%, 8th of 50, and Kimi K3 at 23.0% ± 3.7%. How to read gaps like that is in LLM benchmarks and leaderboards. For K3.1 Pro against DeepSeek and Qwen, with Claude Code, Cursor, and Ollama setup, see DeepSeek vs Qwen vs Kimi for coding.
Is there a Kimi K3.1 beyond K3.1 Pro, or a K2.x still coming?
Moonshot's pages list one K3.1 model, Kimi K3.1 Pro, with its Highspeed variant, and call it "the first model in the K3.1 series." Moonshot's model list, blog, and Hugging Face page name no other K3.1 model and no K4. The K2 line ends at K2.7 Code, and Moonshot's stated plan for it is retirement by the end of December 2026.
Build and publish a page with Kimi
Kimi K3.1 Pro writes a complete web page from one prompt, and Kimi's blog shows front-end pages and 3D scenes it built end to end. When Kimi hands you a page as one HTML file, drop it into Convika's publish tool and click Get my preview link for a private link that works for 3 minutes, with no account. Publish it from a free account to keep it at one address, with a form that saves what visitors send.
Frequently asked questions
What is the latest Kimi model?
Kimi K3.1 Pro, dated October 24, 2026 on Moonshot AI's blog, with weights on Hugging Face since early November. It has 2.8T total parameters, a 1,048,576-token context, and costs $2 / $10 per million tokens on the Kimi API. Read on November 18, 2026.
When was Kimi K3 released?
Moonshot's blog dates Kimi K3 July 16, 2026, when it became available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The post promised the full weights "by July 27, 2026." Kimi K3.1 Pro followed on October 24, 2026.
How much does Kimi K3 cost?
On the Kimi API, kimi-k3 costs $3 per million input tokens and $15 per million output tokens, and kimi-k3.1-pro costs $2 / $10. The API has no subscription; K3 unlocks after a top-up of at least $1. The Kimi app's membership starts at Plus, $19 a month.
Is Kimi K3 free?
The weights are free to download from Hugging Face under the Kimi K3 License. Kimi's plans page lists a Free plan at $0 with 1 agent task at a time. The API is paid by the token.
Is Kimi K3 open source?
Kimi K3 and K3.1 Pro have open weights on Hugging Face under Moonshot's own licenses. They allow commercial use, with two conditions: products above 100 million monthly users or $20 million monthly revenue must display the model's name, and Model-as-a-Service businesses above $20 million in 12-month revenue need a separate agreement. The K2 line uses a Modified MIT License with the display rule only.
How big is Kimi K3?
2.8 trillion parameters in total, with 104 billion activated per token. Its weight files on Hugging Face total 1,561 GB. The K2.5, K2.6, and K2.7 Code repositories are 595 GB each, for 1T-parameter models with 32B activated.
What is the difference between Kimi K2.5, K2.6, and K3?
K2.5 (January 2026) added vision to K2. K2.6 (April 2026) kept the same 1T-parameter architecture and improved long-horizon coding. K3 (July 2026) is a new 2.8T-parameter model with a 1,048,576-token context, against 262,144, plus video input. K2.5 is retired from the API, and K2.6 closes in late December 2026.
Can I still use Kimi K2 or K2.5?
Not on the Kimi API. Moonshot retired the kimi-k2 series in May 2026 and kimi-k2.5 in August 2026; calls to K2.5 "return a 404". K2.6 and K2.7 Code work for users registered before November 1, with deprecation expected in late December 2026. The weights of all of them remain on Hugging Face.
What is Kimi K2.7 Code?
A coding model from June 2026, "built upon Kimi K2.6", that uses about 30% fewer thinking tokens than K2.6. It costs $0.95 / $4 per million tokens for existing users and retires with the rest of the K2 line.
Sources
Each page was read on November 18, 2026, between 03:35 and 04:05 Japan time; dates in brackets are the page's own.
- Moonshot's blog: Research index, Kimi K3.1 Pro (October 24, 2026), Kimi K3 (July 16, 2026)
- Kimi API platform: Model list, Model pricing, Platform changelog, Kimi K3 quickstart, Kimi K3.1 Pro quickstart
- Hugging Face: Moonshot AI's models, Kimi-K3.1-Pro and its license, Kimi-K3 and its license, Kimi-K2.7-Code, Kimi-K2.6 and its license, Kimi-K2.5, Kimi-K2-Thinking, Kimi-K2-Instruct-0905, Kimi-K2-Instruct
- Plans: Kimi membership pricing
- Independent leaderboard: Terminal-Bench
Next step
How to password protect an HTML page or website
A JavaScript password script leaves the password in the page source. To password protect an HTML page or link, encrypt the file or let a server check it.
Related articles
- DeepSeek vs Qwen vs Kimi for coding (November 2026): latest models, benchmarks, API prices, and where to run themDeepSeek V4.1 Flash, Qwen4-Max, and Kimi K3.1 Pro for coding: each lab's own benchmarks, API and OpenRouter prices, Claude Code, Cursor, Cline, and Ollama.
- Chinese AI models in November 2026: the latest from DeepSeek, Qwen, Kimi, GLM, and MiniMaxEvery major Chinese AI model, from DeepSeek-V4.1 and Qwen4-Max to Kimi K3.1 Pro, GLM-5.3, and MiniMax M3.1: open weights, license, API price, and where to use them.
- The best AI models right now (November 2026): GPT-6, Claude, Gemini, and Grok comparedGPT-6 Astra, Claude Fable 5.2 and Opus 5.5, Gemini 4 Carbon, and Grok 4.7 compared: release dates, context, API prices, plans, and each lab's own benchmarks.
- We gave 9 AI models the same landing page prompt: here's what they builtOne landing page prompt, sent once to GPT-6 Astra, Claude, Gemini 4, Grok 4.7, DeepSeek, Qwen, Kimi and GLM on Nov 16, 2026. Cost, time, phone test.
- The best AI model for a one-page HTML website (November 2026), and how to put the page onlineClaude Opus 5.5 and Fable 5.2 top Arena's WebDev board; Opus built our test page in 57 seconds for $0.15. Put it online: drop the .html file on Convika.
- GPT-6 vs Claude Opus 5.5 (November 2026): which is better for coding, and what each costsGPT-6 Astra vs Claude Opus 5.5 for coding: each lab's own benchmarks, API prices with long-prompt rates, plans, GPT-6 vs GPT-5.6, and Astra vs Fable 5.2.
Skipbox
Cheap & fast GitHub Actions runners, unlimited for $39 a month
Your own machine keeps Docker layers and package caches between jobs. Switch by changing one runs-on line.
Start a free Skipbox trial