Kimi K3
Kimi K3 is Moonshot AI's largest model, released on Jul 16, 2026, with 2.8 trillion parameters, 104 billion active per token, and open weights on Hugging Face since Jul 27, 2026. On the official API it costs US$ 3 per million input tokens and US$ 15 per million output, with a 1,048,576 token window. Moonshot itself says it still trails Claude Fable 5 and GPT-5.6 Sol; its edge is running on your own infrastructure, with the data in-house.
Kimi K3 in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Moonshot AI
- API identifier
kimi-k3- Release
- Jul 16, 2026; weights published on Jul 27, 2026
- Weights
- open on Hugging Face, under the Kimi K3 License
- Size
- 2.8 trillion parameters, 104 billion active (16 of 896 experts per token)
- Input
- US$ 3 per million tokens
- Output (includes reasoning)
- US$ 15 per million tokens
- Cache read
- US$ 0.30 per million
- Cache write
- US$ 3 (5 minutes) or US$ 6 (1 hour)
- Batch API
- not available for K3
- Context window
- 1,048,576 tokens, single price
- Max output
- 131,072 tokens by default; accepts up to 1,048,576 if input and output fit in the window
- Reasoning
- always on; effort low, high or max (default max)
- Input and output
- text, image and video in; text out
- Reliable knowledge
- date not published by Moonshot
- Where it runs
- Moonshot API, Kimi app, Kimi Work, Kimi Code and, with the weights, on vLLM, SGLang or TokenSpeed
Where Kimi K3 does well, where it falls behind and who it is for
Where it does well
- Long coding work: 42.0 on SWE-Marathon and 77.8 on ProgramBench, the top scores in Moonshot's table.
- Web research: 91.2 on BrowseComp and 95.0 on DeepSearchQA, leading both rows of the same table.
- Tools and legal work: 94.5 on MCPMark-Verified and 94.6 on Harvey Lab-AA, also at the top.
- Open weights: you can download it, run it on your own servers and fine-tune it on company data.
- Large context at a single price: 1,048,576 tokens with no pricier tier; GPT-6.1 Sol charges double for input above 272 thousand.
Where it falls behind
- Behind the closed leaders: 67.5 on DeepSWE against 73.0 for GPT-5.6 Sol, and 58.3 on OSWorld 2.0 against 66.1 for Claude Fable 5, in Moonshot's own figures.
- API pricier than closed peers: US$ 3 and US$ 15, against US$ 2 and US$ 10 for Claude Sonnet 5.5 and GPT-6.1 Sol. Cache reads cost US$ 0.30, three times GPT-6.1 Sol.
- Demanding on the agent: it needs all previous reasoning passed back. Without that, or when switching models mid-session, Moonshot warns that quality may become highly unstable.
- Too proactive: it may make unexpected decisions on the user's behalf, and Moonshot admits a noticeable gap in user experience against Fable 5 and GPT-5.6 Sol.
Who it makes sense for
- Companies that cannot send sensitive data to a third-party API and own, or rent, large-scale GPU infrastructure.
- Long-running coding agents on Kimi Code, the harness Moonshot recommends.
- Deep web research with a sourced report, where it leads BrowseComp and DeepSearchQA in Moonshot's table.
- Anyone who does not want to depend on a vendor's schedule: Moonshot retired Kimi K2.5 from its API on Aug 31, 2026, and the weights stay with you.
When to pick another model
- Low volume and no infrastructure team: Moonshot recommends serving K3 on supernodes with 64 or more accelerators.
- Sensitive data on the official API: the terms let Moonshot use customer content to improve its services unless agreed otherwise in writing.
- Agents that read email or web pages without their own safeguards: in a Google chart with Gray Swan data, prompt injection attacks within 15 attempts succeeded 52.7% of the time against K3, against 1.0% for Claude Opus 5.5.
- Short, instant answers: K3 always reasons, and the lowest accepted effort is low.
What Moonshot published at launch
K3 came out on Jul 16, 2026 in the Kimi app, Kimi Work, Kimi Code and the API, and the weights reached Hugging Face on Jul 27, 2026, as promised. It is a mixture-of-experts model: only 16 of 896 experts work on each token, which gives 104 billion active parameters out of 2.8 trillion. Moonshot says the new architecture delivers about 2.5 times the scaling efficiency of Kimi K2.
The launch text is candid: overall, K3 still trails Claude Fable 5 and GPT-5.6 Sol. In the 45-test table it leads on long coding work, web research, tools and legal work, and loses to Fable 5 on most of the rest.
Two caveats. The rivals in the table already have newer versions: Fable 5.1, Opus 5.5, GPT-6 Astra and GPT-6.1 Sol came out between Sep 1 and Sep 29, 2026 and are not in it. And several tests used Moonshot's own harness.
| Kimi K3 Moonshot | Fable 5 Anthropic | GPT-5.6 Sol OpenAI | Opus 4.8 Anthropic | GLM-5.2 Z.ai | |
|---|---|---|---|---|---|
| SWE-Marathon long coding work |
42.0 | 35.0 | 39.0 | 40.0 | 13.0 |
| ProgramBench coding |
77.8 | 76.8 | 77.6 | 71.9 | 63.7 |
| DeepSWE v1.1 long-horizon coding |
67.5 | 70.0 | 73.0 | 59.0 | 46.2 |
| FrontierSWE coding |
81.2 | 86.6 | 71.3 | 66.7 | 67.3 |
| Terminal-Bench 2.1 terminal |
88.3 | 88.0 | 88.8 | 84.6 | 82.7 |
| BrowseComp web research |
91.2 | 88.0 | 90.4 | 84.3 | not published |
| MCPMark-Verified tool use |
94.5 | 87.4 | 92.9 | 76.4 | not published |
| GDPval-AA v2 office work, Elo |
1,686 | 1,747 | 1,736 | 1,593 | 1,510 |
| OSWorld 2.0 computer use |
58.3 | 66.1 | 62.6 | 55.7 | not published |
| Harvey Lab-AA legal work |
94.6 | 93.6 | 87.2 | 91.1 | 91.0 |
| HLE-Full hard knowledge, without and with tools |
43.5 and 56.0 | 53.3 and 63.0 | 44.5 and 58.0 | 49.8 and 57.9 | not published |
| GPQA Diamond science |
93.5 | 92.6 | 94.1 | 91.0 | 91.2 |
Figures from the Kimi K3 card on Hugging Face, as Moonshot published them, without the percent sign; all at maximum effort. The original table also includes GPT-5.5. On SWE-Marathon, Fable 5 hit fallbacks on 35% of the tasks, according to Moonshot. The dot marks the best in each row.
Open weights: what changes for a company
Open weights means the whole model can be downloaded and run on your own or rented servers. Three things change: prompts never leave your infrastructure, the model does not disappear if the API changes its price or schedule, and fine-tuning on company data stays with you.
The cost becomes infrastructure: the weights add up to about 1.56 TB, and Moonshot recommends serving K3 on supernodes with 64 or more accelerators, with vLLM, SGLang or TokenSpeed.
The official API is different: the terms, updated on Jul 30, 2026, let Moonshot use customer content to improve its services unless agreed otherwise in writing, and data is kept on servers in Singapore.
The Kimi K3 License allows commercial use and sets two conditions that only reach very large operations:
- You may use, modify, fine-tune, distribute and sell it for free, keeping the copyright notice, with no warranty of any kind.
- Anyone selling the model as a service, giving third parties access to inference or fine-tuning, with revenue above US$ 20 million over 12 months needs a separate agreement with Moonshot.
- Commercial products with more than 100 million monthly active users, or more than US$ 20 million in monthly revenue, must prominently display "Kimi K3".
- Internal use, which does not hand the model or its outputs to third parties, falls outside both conditions.
Pricing: K3 on the API against closed models
Through the API, K3 is not the cheapest in its tier: it costs 50% more than Claude Sonnet 5.5 and GPT-6.1 Sol, and 25% less than Claude Opus 5.5. Moonshot says it hits the cache on more than 90% of coding calls, which would keep most input at US$ 0.30. No maker has put the four in the same table; the independent scores are at the top of this page.
| Kimi K3 Moonshot | Sonnet 5.5 Anthropic | GPT-6.1 Sol OpenAI | Opus 5.5 Anthropic | |
|---|---|---|---|---|
| Input | US$ 3 | US$ 2 | US$ 2 | US$ 4 |
| Output | US$ 15 | US$ 10 | US$ 10 | US$ 20 |
| Cache read | US$ 0.30 | US$ 0.20 | US$ 0.10 | US$ 0.20 |
| Context window | 1,048,576 | 1 million | 1,050,000 | 1 million |
| Long prompts | single price | single price | double input above 272 thousand | single price |
| Downloadable weights | yes | no | no | no |
Standard pricing. Moonshot, Anthropic and OpenAI documentation checked on Oct 1, 2026.
How to test before adopting
Start with the API, with ten real tasks on K3 and on your current model, comparing accuracy and cost per task. Use a harness that passes the full reasoning back, such as Kimi Code, and set clear boundaries in the system prompt. Also test high and low effort, which spend less than the default max. If the result holds up and the data cannot leave the building, then price your own servers.
Examples of using Kimi K3
K3 runs on the company's servers with vLLM or SGLang and works in the repository through Kimi Code. It is internal use, outside the license conditions, and it is the kind of task where it leads SWE-Marathon in Moonshot's table, with 42.0.
An agent searches the web, cross-checks dozens of pages and delivers a report with links. K3 scores 91.2 on BrowseComp and 95.0 on DeepSearchQA in Moonshot's figures.
Contracts with customer data go through the model without leaving the company network. On Harvey Lab-AA, a legal work test, Moonshot published 94.6 for K3, the top score in the table.
Where Kimi K3 stands out and where it falls behind, by the numbers
Kimi K3 has published results on 30 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Oct 19, 2025 and Oct 14, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 30 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 93.1% | #19 / 80 | #19 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 97.2% | #30 / 82 (tied with 1) | #31 / 168 | 100.0% GPT-5.5 |
| Dtbench | 91.2% | #30 / 69 (tied with 1) | #30 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 52.74 | #17 / 69 | #17 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 23.4% | #19 / 79 | #19 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 82.6% | #9 / 54 | #9 / 112 | 93.6% GPT-6 Astra |
| SciCode | 59.5% | #10 / 76 | #10 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 72.2% | #24 / 67 | #24 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 60.4% | #29 / 50 (tied with 1) | #29 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 1,524 | #14 / 64 | #14 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,674 | #9 / 75 | #9 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 94.5% | #15 / 50 (tied with 2) | #15 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 50.6% | #25 / 63 | #26 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 36.0% | #37 / 64 (tied with 1) | #39 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 26.0% | #32 / 62 (tied with 1) | #33 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 87.0% | #10 / 70 | #10 / 74 | 100.0% Claude Fable 5.1 |
| SimpleBench | 60.7% | #21 / 38 | #23 / 74 | 81.9% Claude Fable 5 |
| Frontiermath Tier 4 V2 | 39.0% | #24 / 57 | #24 / 67 | 100.0% GPT-6.1 Sol |
| ForecastBench | 61.90 | #1 / 21 | #3 / 62 | 62.50 o3 |
| Vending Bench 2 | 5,165 | #28 / 49 | #28 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 50.6% | #26 / 49 | #26 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 44.2% | #16 / 43 | #16 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 19.0% | #24 / 37 | #24 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 34.2% | #21 / 34 (tied with 1) | #21 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 29.5% | #18 / 30 | #18 / 31 | 54.4% Gemini 4 Argon |
| DeepSWE | 68.5% | #8 / 26 | #8 / 26 | 74.1% GPT-6 Astra |
| Surface Evolver Bench | 95.0% | #1 / 23 (tied with 1) | #1 / 24 | 95.0% Claude Fable 5 |
| Gbaeval | 48.3% | #9 / 22 | #9 / 22 | 79.6% Claude Opus 5 |
| FrontierSWE | 25.9% | #12 / 20 | #12 / 20 | 65.5% GPT-6 Astra |
| Posttrainbench | 32.0% | #5 / 11 | #5 / 11 | 41.8% Claude Fable 5 |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Kimi K3 with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Gemini 3.7 Flash | Google DeepMind | 157.22 | 1,551 | 78.8 | Aug 13, 2026 |
| Gemini 3.8 Flash | Google DeepMind | 157.20 | 1,551 | 75.8 | Sep 2, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| GPT-5.3 Codex | OpenAI | 156.76 | n/a | n/a | Feb 5, 2026 |
| GPT-5.4 | OpenAI | 156.72 | 1,519 | 78.0 | Mar 5, 2026 |
Kimi K3 or Gemini 3.7 Flash?
Gemini 3.7 Flash is the model from another lab closest to Kimi K3 in capability. Across the 22 tests both took, this is the score:
| Test | Kimi K3 | Gemini 3.7 Flash |
|---|---|---|
| Chess puzzles | 36.0% | 66.0% |
| ALE-Bench | 1,524 | 904 |
| CritPt | 23.4% | 14.3% |
| ProofBench | 87.0% | 58.0% |
| Game puzzles | 26.0% | 37.0% |
| ARC-AGI 2 | 60.4% | 84.6% |
| SimpleQA Verified | 50.6% | 69.2% |
| Apex Agents | 50.6% | 67.8% |
| Furniture Assembly | 34.2% | 26.7% |
| FrontierSWE | 25.9% | 20.3% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Kimi K3 costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
ARC-AGI 2
The score scale does not start at zero.
ARC-AGI
The score scale does not start at zero.
Other models from Moonshot
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Kimi K3 | Jul 16, 2026 | 157.42 | 1,528 | 79.2 | Open weights (non-commercial) |
| Kimi K2.7 Code | Jun 12, 2026 | 149.87 | n/a | 68.4 | Open weights (unrestricted) |
| Kimi K2.6 | Apr 20, 2026 | 150.92 | 1,521 | 70.5 | Open weights (unrestricted) |
| Kimi K2.5 | Jan 27, 2026 | 147.98 | 1,495 | n/a | Open weights (unrestricted) |
| Kimi K2 Thinking | Nov 6, 2025 | 145.96 | n/a | n/a | Open weights (restricted use) |
| Kimi K2 Thinking Turbo | Nov 6, 2025 | 145.79 | 1,456 | n/a | Open weights (restricted use) |
| Kimi K2 Instruct | Jul 12, 2025 | 140.13 | n/a | n/a | Open weights (restricted use) |
Have you used Kimi K3? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Kimi K3
How much does Kimi K3 cost?
On the Moonshot API, US$ 3 per million input tokens, US$ 15 per million output and US$ 0.30 per million for cache reads. Writing to the cache costs US$ 3 (5 minutes) or US$ 6 (1 hour), and there is no batch discount for K3. The weights are free, but running the model takes large infrastructure.
Is Kimi K3 open source?
The weights and code were released under the Kimi K3 License, which allows using, modifying, fine-tuning and selling without paying Moonshot. Only very large operations face conditions: a separate agreement for selling the model as a service above US$ 20 million over 12 months, and the Kimi K3 name on screen for products with more than 100 million monthly users or US$ 20 million in monthly revenue. Internal use falls outside both.
Can I run Kimi K3 on my own servers?
Yes, but it is not small. The weights add up to about 1.56 TB, and Moonshot recommends supernodes with 64 or more accelerators, running vLLM, SGLang or TokenSpeed.
Is Kimi K3 better than Claude and GPT?
Moonshot itself says that, overall, it still trails Claude Fable 5 and GPT-5.6 Sol, although it leads tests in long coding work, web research and legal work. Newer models, such as Claude Opus 5.5 and GPT-6.1 Sol, are not in the comparison; the independent scores are at the top of this page.
Is my data safe on the Moonshot API?
The terms let Moonshot use submitted content to improve its services unless agreed otherwise in writing, and data is kept on servers in Singapore. For sensitive data, the way out is running the weights on your own infrastructure.
Where does Kimi K3 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Kimi K3 ranks #17 out of 267 models, with an ECI of 157.42. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Kimi K3 better than Gemini 3.7 Flash?
Across the 22 tests both models took and Epoch AI compiled, Kimi K3 comes out ahead on 11, behind on 10 and tied on 1. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Kimi K3 best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Kimi K3 are: ForecastBench (#1 of 21); Surface Evolver Bench (#1 of 23); WebDev Arena (#9 of 75). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Moonshot AI, Kimi K3 tech blog
- Moonshot AI, Kimi K3 card on Hugging Face (architecture and benchmarks)
- Moonshot AI, Kimi K3 License
- Moonshot AI, Kimi K3 technical report (PDF)
- Moonshot AI, home page (release date)
- Kimi API, pricing
- Kimi API, batch pricing
- Kimi API, Kimi K3 quickstart
- Kimi API, model parameter reference
- Kimi API, Chat Completions reference (max output)
- Kimi API, model list and retirements
- Kimi API, terms of service
- Kimi API, privacy policy
- Google, Gemini 4 Argon announcement (Gray Swan prompt injection chart)
- Anthropic, pricing (documentation)
- OpenAI, API pricing
- OpenAI, GPT-6.1 Sol page (documentation)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking