Kimi K3 has 2.8T parameters, but only 104B run per token
Kimi K3 is the most capable model from Moonshot AI so far, and Moonshot calls it the world's first open 3T-class model. It is available in the Kimi apps, in Kimi Code and through the Kimi API. The full weights went up on Hugging Face on July 27, 2026.
The 2.8T figure is easy to misread: it is the total parameter count, but each token passes through only 104B parameters, under 4% of the total. API prices scale with the tokens you use, while hosting the model means loading all 2.8T parameters.
Kimi K3 at a glance (Hugging Face model card)
Only 16 of the 896 experts are chosen for each token
MoE (Mixture-of-Experts) is a design that holds many specialized sub-networks, called experts, and runs only a few of them for each input. Kimi K3 has 896 experts and picks 16 for every token, plus 2 shared experts that always run. That selection is why 104B parameters, not 2.8T, do the work on each token.
The model has 93 layers: 69 use Kimi Delta Attention and 24 use Gated MLA. Moonshot credits this architecture with roughly 2.5 times the overall scaling efficiency of Kimi K2.
Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), and scales up MoE sparsity with a Stable LatentMoE framework that activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2. ／ Total Parameters ／ 2.8T ／ Activated Parameters ／ 104B ／ Number of Shared Experts ／ 2 ／ Attention-Layer Composition ／ 69 KDA + 24 Gated MLA — From the Key Features list and the cells of the Model Summary tableView official source →
A 1M-token context window, plus image and video input
The context length is 1,048,576 tokens, enough to pass a long specification or a large part of a codebase in one request. A token is the unit a model uses to count text, often a word or part of a word.
Vision was built in rather than added later. The model card says K3 understands text, images and video within the same model, but its summary table lists only text and image as modalities.
Kimi K3 understands text, images, and video within the same model, and supports a 1-million-token context window. ／ Context Length ／ 1048576 ／ Modality ／ Text, Image — From the Native Multimodality & Long Context feature and the Model Summary tableView official source →
Kimi K3 benchmark scores depend on the harness behind them
A benchmark is a fixed set of tasks every model is scored on. A harness is the agent program that runs the model through those tasks, such as Kimi Code, Claude Code or Codex. Moonshot didn't test every model in the same harness, so no single number settles whether K3 beats Claude.
| Test in the model card | Kimi K3 | Fable 5 | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|
| FrontierSWE | 81.2 | 86.6 | 66.7 | 71.3 |
| DeepSWE | 67.5 | 70.0 | 59.0 | 73.0 |
| SWE-Marathon | 42.0 | 35.0 | 40.0 | 39.0 |
| Terminal-Bench 2.1 | 88.3 | 88.0 | 84.6 | 88.8 |
| GPQA Diamond | 93.5 | 92.6 | 91.0 | 94.1 |
The table is a selection from Moonshot's model card, with Kimi K3 at its maximum reasoning setting. Moonshot's footnotes explain how each score was produced, and on SWE-Marathon they change how the row reads.
Moonshot says it still trails Claude Fable 5 and GPT-5.6 Sol
Moonshot's launch post opens with a caveat of its own: overall, Kimi K3 still trails Claude Fable 5 and GPT-5.6 Sol, the strongest proprietary models. It then says K3 performed at frontier level across its evaluation suite and consistently beat the other models Moonshot tested.
The same post lists three limitations. Output quality can become highly unstable if a harness doesn't pass back the model's earlier thinking, and K3 may make unexpected decisions when instructions are unclear. Moonshot also admits a noticeable gap in user experience compared with Fable 5 and GPT-5.6 Sol.
While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, Kimi K3 demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models. ／ generation quality may become highly unstable ／ it may make unexpected decisions on the user's behalf ／ K3 nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol. — From the opening of the launch post and its Limitations sectionView official source →
Coding winners change with the harness and with fallbacks
Kimi K3 leads some coding benchmarks and trails on others. It scores 88.3 on Terminal-Bench 2.1, just above Fable 5 at 88.0 and below GPT-5.6 Sol at 88.8. On FrontierSWE it scores 81.2 against Fable 5's 86.6.
The test conditions also differ by model. Every Kimi K3 score uses maximum reasoning effort and a temperature of 1.0, and on Terminal-Bench 2.1, K3 ran in Kimi Code while the other models got their best score across harnesses. On DeepSWE, K3 scores 67.5 in Kimi Code and 67.3 in the mini-SWE-agent harness.
SWE-Marathon needs the most care. K3's 42.0 beats Fable 5's 35.0, but Moonshot's footnote says Fable 5 hit fallbacks on 35% of those tasks, without explaining what triggered them, and that this may have lowered its measured score.
All Kimi K3 results are obtained with reasoning effort set to 'max' and temperature = 1.0. ／ For all other models, we report the best score across harnesses ／ under which Kimi K3 attains 67.3 with the mini-SWE-agent harness ／ Additionally, Claude Fable 5 hit fallbacks on 35% of the tasks in our evaluation, which may have negatively impacted its measured performance. ／ Terminal-Bench 2.1 ／ 88.3 ／ 88.0 ／ 88.8 ／ FrontierSWE ／ 81.2 ／ 86.6 ／ SWE-Marathon ／ 42.0 ／ 35.0 ／ DeepSWE ／ 67.5 ／ 70.0 ／ 73.0 ／ GPQA Diamond ／ 93.5 ／ 92.6 ／ 94.1 — From the table footnotes and the scores in the Evaluation Results tableView official source →
API access is cheaper than serving the 1.56 TB open weights
There are three ways to use Kimi K3: the Kimi apps, the Kimi API, or the open weights on your own servers. The apps run on iOS, Android and HarmonyOS, and developers can select kimi-k3 on the Kimi API Platform or switch to it with /model in Kimi Code. For most teams the API is the realistic route, because self-hosting requires a GPU cluster.
Ways to use Kimi K3 (checked September 15, 2026)
API pricing is $3 input and $15 output per million tokens
The Kimi API charges $3.00 per million input tokens, $0.30 when the input hits the cache, and $15.00 per million output tokens. A cache hit means that part of the prompt was already processed, so repeated long context costs 90% less. Moonshot says its API hits the cache more than 90% of the time on coding workloads.
At Anthropic's list prices, Kimi K3's API costs 30% of Claude Fable 5's $10 input and $50 output per million tokens. It matches Sonnet 4.6 at $3 and $15, and Sonnet 5 is lower still at $2 and $10. The rest of Anthropic's lineup is in Claude models compared: Fable, Opus, Sonnet, Haiku.
Pricing is $0.30/MTok for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output. Powered by Mooncake's disaggregated inference architecture, the official Kimi API achieves a cache hit rate above 90% in coding workloads. — From the Build with the Kimi API item in the Availability sectionView official source →
Claude Fable 5 ／ $10 / MTok ／ $50 / MTok ／ Claude Sonnet 5 ／ $2 / MTok ／ $10 / MTok ／ Claude Sonnet 4.6 ／ $3 / MTok ／ $15 / MTok — From the base input and output columns of the model pricing tableView official source →
Self-hosting takes 1.56 TB of weights and 64+ accelerators
Open weights means anyone can download them, not that anyone can run them. The repository holds 96 safetensors files totaling about 1.56 TB, according to a Hugging Face API count on September 15, 2026. Only 104B parameters run per token, but the full set of weights still has to be loaded to serve requests.
Moonshot recommends supernode configurations with 64 or more accelerators, because inference runs more efficiently when more accelerators share one high-bandwidth interconnect. It names vLLM, SGLang and TokenSpeed as the inference engines to use.
That makes K3 a cluster project. For experiments on your own machine, a smaller model is a better fit, as Local LLM Coding Now Hits 75% of Cloud AI explains. Another large open-weight release is covered in GLM 5.2 benchmarks, pricing, and Claude Code setup.
The low-precision weights were planned from training, not applied afterward. Moonshot trained K3 with 4-bit MXFP4 weights and 8-bit MXFP8 activations from the SFT stage on, so the files on Hugging Face are the finished model.
Kimi K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility. ／ Since inference efficiency likewise benefits from larger high-bandwidth communication domains, we recommend deploying Kimi K3 on supernode configurations with 64 or more accelerators. — From the statements on quantization and on the recommended deploymentView official source →
Currently, Kimi K3 is recommended to run on the following inference engines: ／ vLLM ／ SGLang ／ TokenSpeed — From the list of recommended inference enginesView official source →
The Kimi K3 License adds terms only for very large companies
Both the code and the weights come under the Kimi K3 License, Moonshot AI's own license rather than MIT or Apache. Its opening grant is broad: free of charge, you may use, copy, modify, distribute, sell, deploy and fine-tune the software. Two business conditions sit on top of that.
The first covers Model as a Service, meaning a business that gives third parties API-style access to inference or fine-tuning. Once such a business and its affiliates pass $20 million in revenue in any 12 consecutive months, they must sign a separate agreement with Moonshot AI before any commercial use. The second covers any product past 100 million monthly active users or $20 million in revenue in a single month: it has to show the Kimi K3 name prominently on screen.
Neither condition applies to internal use, where no outputs or capabilities reach third parties, or to use through Moonshot's official products and certified inference partners. A company running K3 only for its own staff falls outside both.
exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose. ／ that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue ／ must be prominently displayed on the user interface of such product or service ／ internal use of the Software, defined as any use that does not make the Software, its outputs, or its underlying capabilities available to third parties — From the Model as a Service, attribution and exemption sectionsView official source →
Summary: use the Kimi K3 API unless you already run a GPU cluster
Kimi K3 is a 2.8T-parameter MoE model, but only 104B parameters run for each token. Moonshot itself places it behind Claude Fable 5 and GPT-5.6 Sol overall, and which model wins on coding shifts with the harness and with Fable 5's fallbacks.
For most readers the Kimi K3 API is the way in, at $3 input and $15 output per million tokens, 30% of Fable 5's price. The 1.56 TB of weights suits teams that already run 64 or more accelerators, and the license only adds terms for businesses past $20 million in revenue (over 12 months for API services, in one month for products) or 100 million monthly active users. Model cards and license files like these are long pages full of tables.
FAQ
- Is Kimi K3 open source?
- The weights are free to download from Hugging Face, but they are released under Moonshot AI's own Kimi K3 License rather than a standard license such as MIT or Apache. Open-weight is the more accurate label, and the license terms are worth reading before commercial use.
- Can I run Kimi K3 locally?
- Not on a personal computer. The 96 weight files total about 1.56 TB, and Moonshot AI recommends supernode configurations with 64 or more accelerators. Most people will use the API or the Kimi apps instead.
- How much does the Kimi K3 API cost?
- The Kimi API charges $3.00 per million input tokens, $0.30 when the input hits the cache, and $15.00 per million output tokens. That is 30% of Claude Fable 5's list price of $10 input and $50 output.
- Is Kimi K3 better than Claude Fable 5?
- Not overall, by Moonshot AI's own account. It edges ahead of Fable 5 on some coding benchmarks, such as Terminal-Bench 2.1 at 88.3 against 88.0, and falls behind on others, such as FrontierSWE at 81.2 against 86.6.
Related articles
Kimi K3.1 Pro benchmarks, pricing and open weightsKimi K3.1 Pro is Moonshot AI's post-trained Kimi K3. It nearly matches Claude Opus 5.5 on coding, costs a third less than K3 and now has open weights.
Kimi K3 Distillation: What the White House ClaimsThe Kimi K3 distillation claim: the White House says Moonshot AI distilled Anthropic's Fable. What Kratsios posted, what Moonshot denies, what is unproven.
Chinese Open Source AI Faces Possible US CurbsChinese open source AI models may face new US restrictions. The trigger, Kimi K3, the measures under discussion, and the split inside the administration.
Mistral Large 4 (Le Chonk) specs and benchmarksMistral Large 4, a 1.05T-parameter MoE with 52B active, went open-weight on October 29 under a modified MIT license. Input now costs $0.70 per million tokens.