Text and chat · Moonshot AI
Kimi
Sub Rosa's suggested alternate: an open model that excels at long tool-driven tasks and reads images, in private mode.
Private Moderate costWhat sets it apart
Kimi is built for long agent runs that chain many tool calls, and its open models are multimodal from K2.5 on (text, image, video in). K3 is the largest open model in the catalog at 2.8T parameters, with a 1M context window.
- 1T MoE, then 2.8T with K3
- Reads images and video
- Open weights, modified licenses
Recommended for: think through a hard problem or a long task ; write or fix code ; ask about a photo, a screenshot or a scanned page ; keep a sensitive conversation as private as possible.
Strengths and limits
Where it shines
- Sub Rosa switches to K2.6 when you attach an image
- Ranked the leading open-weight model by Artificial Analysis at release
- Excels at long tasks that chain many tool calls
- Private mode, and a version listed as end-to-end encrypted
Think twice when
- K3 and K3 Fast are among the dearer open models per token
- K2.6 context stops at 256K tokens
- K2.6: 256K context, vision, tool calling
- K3: 1M context, several images per message
- Open weights
Use cases
A long research task with many steps
Search my notes for every mention of the Lisbon supplier, then draft a one-page timeline of what was promised and when, and flag anything that conflicts.
Kimi is tuned for tasks that chain many tool calls in a row.
Describing a chart or screenshot
This is a screenshot of our sales dashboard. Which region fell the most month over month, and what would you check first?
K2.6 and K3 accept images; Sub Rosa switches to K2.6 when you attach one.
Writing and reviewing code
Write a Python script that renames every photo in a folder to its capture date, then explain how to undo it.
K2.7 Code is the coding-focused version and always reasons before answering.
Against its rivals
- GLMGLM 5.3 scores a point higher on the index (45 vs 44) at a lower token price.#13 of 90 on Artificial Analysis Intelligence Index (45 points)
- DeepSeekDeepSeek costs far less per token; Kimi leads on the index and reads images.#24 of 90 on Artificial Analysis Intelligence Index (39 points)
- Qwen open modelsQwen’s open models come in many sizes; Kimi K3 is a single very large model.#23 of 90 on Artificial Analysis Intelligence Index (40 points)
Version history
Each version, dated, with what it changed over the one before.
Hollow points are earlier versions no longer offered in Sub Rosa.
See the numbers
| Version | Released | points |
|---|---|---|
| Kimi K2 Thinking | 2025-11-06 | 22 |
| Kimi K2.5 | 2026-01-27 | 23 |
| Kimi K2.6 | 2026-04-20 | 27 |
| Kimi K2.7 Code | 2026-06-12 | 26 |
| Kimi K3 | 2026-07-16 | 44 |
- Kimi K2 ThinkingEarlier, not in Sub Rosa
- Introduced the 1T-parameter MoE (32B active) with a 256K context, text only, as a reasoning model.
- Kimi K2.5
- Adds image and video input to the same 1T MoE, still 256K context.
- Kimi K2.6
- Same architecture as K2.5 with a higher index score (27 vs 23 on Artificial Analysis).
- Kimi K2.7 Code
- Coding-focused version of the 1T MoE, accepts text, image and video, always reasons.
- Kimi K3Latest
- New architecture with Kimi Delta Attention and Attention Residuals, about 2.5 times K2’s scaling efficiency per the model card.
- 2.8T parameters (104B active) and a 1M token context, up from 1T and 256K.
Benchmarks
How to read these scores: what a benchmark tells you.
Artificial Analysis Intelligence Index: #15 of 90 models measured in the catalog. One score averaging about ten independent tests of reasoning, knowledge, math, coding and agent work, all run the same way for every model. Higher is better. Each model is listed at its highest reasoning effort. Two or three points apart is a close call; ten points is a different class of model.
See the numbers
| Model | points |
|---|---|
| 1. Claude Opus 5.5 | 58 |
| 2. Claude Sonnet 5.5 | 56 |
| 3. Claude Fable 5.1 | 53 |
| 4. GPT-6 Astra | 53 |
| 5. Gemini 4 Argon | 53 |
| 6. GPT-6.1 Sol | 52 |
| 7. Muse Wonder | 52 |
| 8. Claude Opus 5 | 51 |
| 9. Claude Fable 5 | 50 |
| 10. GPT-6 Sol | 48 |
| 15. Kimi K3 | 44 |
| 46. Kimi K2.6 | 27 |
| 49. Kimi K2.7 Code | 26 |
Announced by the maker
Measured by the maker with its own settings: useful, but not directly comparable with the independent scores above.
| Benchmark | Model | Score | Source |
|---|---|---|---|
| gpqa-diamond | Kimi K3 | 93.5 | huggingface.co |
All 7 models in this family
| Model | Use | Privacy | Price |
|---|---|---|---|
Kimi K2.6kimi-k2-6 | Text | Private | $1.574 per million tokens written$0.0011 a page written |
Kimi K2.5kimi-k2-5 | Text | Private | $1.574 per million tokens written$0.0011 a page written |
Kimi K2.7 Codekimi-k2-7-code | Text | Private | $1.574 per million tokens written$0.0011 a page written |
Kimi K3kimi-k3 | Text | Private | $8.43 per million tokens written$0.0059 a page written |
Kimi K3 Fastkimi-k3-fast-api | Text | Private | $10.12 per million tokens written$0.00708 a page written |
Kimi K2.6 E2EEe2ee-kimi-k2-6 | Text | Private | $1.852 per million tokens written$0.0013 a page written |
Kimi K3 E2EEe2ee-kimi-k3-p | Text | Private | $8.43 per million tokens written$0.0059 a page written |
Specifications
| Specification | Kimi K3 | Kimi K2.7 Code | Kimi K2.6 | Kimi K2.5 |
|---|---|---|---|---|
| Released | July 16, 2026 | June 12, 2026 | April 20, 2026 | January 27, 2026 |
| Parameters | 2.8T total, 104B active | 1T total, 32B active | 1T total, 32B active | 1T total, 32B active |
| Weights | Open, Kimi K3 License (commercial use with restrictions) | Open, Modified MIT | Open, Modified MIT | Open, Modified MIT |
| Accepts | text, images | text, images, video | text, images, video | text, images, video |
| Produces | text | text | text | text |
| Context | 1,000K tokens | 256K tokens | 256K tokens | 256K tokens |
| Abilities | Reads images, Uses Sub Rosa’s tools, Thinks before answering, Tuned for code | Reads images, Uses Sub Rosa’s tools, Thinks before answering, Tuned for code | Reads images, Uses Sub Rosa’s tools, Thinks before answering, Tuned for code | Reads images, Uses Sub Rosa’s tools, Thinks before answering, Tuned for code |
| Price | $8.43 per million tokens written | $1.574 per million tokens written | $1.574 per million tokens written | $1.574 per million tokens written |
| Privacy | Private | Private | Private | Private |
Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.
Sources
Everything on this page comes from these pages, read on October 16, 2026 or as dated on each score.
Show the 10 sources
- artificialanalysis.ai/models/kimi-k3
- artificialanalysis.ai/models/kimi-k2-7-code
- artificialanalysis.ai/models/kimi-k2-6
- artificialanalysis.ai/models/kimi-k2-5
- artificialanalysis.ai/models/kimi-k2-thinking
- huggingface.co/moonshotai/Kimi-K3
- artificialanalysis.ai/leaderboards/models
- artificialanalysis.ai/articles/kimi-k2-6-the-new-leading-open-weights-model
- rits.shanghai.nyu.edu/ai/kimi-k3-open-weights-ship-2-8t-parameters-1-4-tb-to-run
- huggingface.co/moonshotai/Kimi-K2.6