othermultimodalRuns locally2.8T
Kimi K3: Specs, License & How to Run
Moonshot 2.8T/104B open-weight multimodal MoE with 1M context under the Kimi K3 License.
- License
- other
- Parameters
- 2.8T (active 104B)
- HF likes
- 11,761 (2026-10-19)
- First seen
- 2026-06-13
TL;DR. Kimi K3 is a multimodal model from Moonshot AI. Parameters: 2.8T (active 104B).License: other. It is listed under ChatGPT, Closed LLM APIs.
| Developer | Moonshot AI |
|---|---|
| Type | multimodal |
| Architecture | Stable LatentMoE; 93 layers (1 dense); 896 experts, 16 selected per token, 2 shared; 69 KDA + 24 Gated MLA; MoonViT-V2 vision encoder (401M); MXFP4 weights / MXFP8 activations |
| Parameters | 2.8T · active 104B |
| Modalities | text, image, video |
| Context | 1048576 |
| License | other — Kimi K3 License (card license_name kimi-k3). Attribution/notice required. If you operate Model-as-a-Service and aggregate revenue exceeds USD 20 million in any consecutive 12 months, enter a separate agreement with Moonshot AI before commercial use of the Software or derivatives. Products/services with more than 100 million MAU or more than USD 20 million monthly revenue must display “Kimi K3” prominently. Sections 2–3 do not apply to internal use or official/certified Moonshot endpoints. Limited commercial terms. Not legal advice. |
| Hardware | Not yet verified — Card: MXFP4 weights / MXFP8 activations via quantization-aware training. HF API safetensors.parameters total 2,779,931,837,184 (~2.78T); usedStorage 1,561,018,243,668 bytes (~1.45 TiB on disk). Peak VRAM is not independently measured. Datacenter-class. (not independently tested) |
| Hugging Face likes | 11,761 (as of 2026-10-19) |
| Model card | https://huggingface.co/moonshotai/Kimi-K3 |
| First seen | 2026-06-13 |
Kimi K3 is an ungated open-weight multimodal model from Moonshot AI. Official hosts we opened on 2026-10-19: moonshotai/Kimi-K3, the Kimi K3 License, github.com/MoonshotAI/Kimi-K3 (8,945 stars), the collection moonshotai/kimi-k3, and platform.kimi.ai Kimi K3 quickstart.
This thin identification page exists so DeepSeek-V4.1-Flash vs Kimi K3 can hang off two sourced entity URLs. It also lists under ChatGPT alternatives and closed LLM APIs. We do not rank writing quality against ChatGPT.
What we verified (2026-10-19)
| Field | Value | Source |
|---|---|---|
| HF likes / downloads | 11,761 / 1,030,720 | HF API |
| Card created / last modified | 2026-06-13T06:42:57Z / 2026-09-02T02:22:41Z | same |
| Gated | false | HF API |
| License (card) | other / kimi-k3 |
HF cardData |
| Pipeline tag | image-text-to-text | HF API |
| Params (README) | 2.8T total / 104B activated | README “Model Summary” |
| safetensors total | 2,779,931,837,184 | HF API |
| usedStorage | 1,561,018,243,668 bytes (~1.45 TiB) | HF API |
| Experts | 896; 16 selected per token; 2 shared | README |
| Context | 1,048,576 | README |
| GitHub stars / forks | 8,945 / 759 | GitHub API |
| GitHub created / last push | 2026-07-27T08:01:37Z / 2026-08-06T10:17:28Z | same |
README introduction also names native vision, video understanding, and a 1M-token window. The model-summary “Modality” row in the extract we opened lists Text, Image; the introduction additionally says the model understands text, images, and video. We copy both and do not reconcile them.
Quantization: MXFP4 weights / MXFP8 activations from SFT onward (card/README). Thinking is always on; reasoning_effort supports low, high, and max (default max). Preserved thinking history must be passed back in messages.
License
LICENSE is titled Kimi K3 License, copyright 2026 Moonshot AI. Hugging Face cardData.license is the opaque label other. We do not invent a more specific SPDX id.
Paraphrase only, not legal advice:
- Keep the copyright notice and permission notice on copies / substantial portions.
- MaaS threshold: if you operate Model-as-a-Service and aggregate revenue of you and affiliates exceeds USD 20 million in any consecutive 12 months, enter a separate agreement with Moonshot AI before using the Software or derivatives for any commercial purpose.
- Display: commercial products/services with more than 100 million MAU or more than USD 20 million monthly revenue must display “Kimi K3” prominently.
- Sections 2 and 3 do not apply to internal use (not making the Software, outputs, or capabilities available to third parties) or to use through Moonshot’s official products or certified inference partners.
- AS IS; no warranty.
Arena.ai snapshot (third-party, not a forkalt ranking)
The Code Arena WebDev leaderboard with the open-source filter, page we opened dated Oct 18, 2026, lists Kimi K3 (Max) at open-source #2 / overall #15, score 1655, 19,426 votes, Moonshot, “Kimi K3 license”, 1M context. DeepSeek-V4.1-Flash-20261017 (Max) sits at open-source #1 / overall #12, score 1670, on the same snapshot. Those are Arena.ai figures. See the comparison.
What we did not verify
README evaluation tables (GPQA Diamond 93.5, DeepSWE 67.5, Terminal-Bench 2.1 88.3, and the rest at reasoning_effort=max) stay self-reported. “World’s first open 3T-class model” is a vendor sentence. We have not loaded MXFP4 on a named cluster.
Related: DeepSeek-V4.1-Flash, Muse Spark Open, Reflection Beam.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| Arena.ai Code Arena WebDev open-source rank (Kimi K3 Max, 2026-10-18 snapshot) | 1655 (#2 open / #15 overall) | https://arena.ai/leaderboard/code/webdev?license=open-source |
| GPQA Diamond (max, vendor) | 93.5 (self-reported) | https://huggingface.co/moonshotai/Kimi-K3 |
Deploy options
- local: Hugging Face weights
- hosted: Kimi platform (kimi-k3)
Comparisons
FAQ
Are Kimi K3 weights public?
Yes. huggingface.co/moonshotai/Kimi-K3 is ungated (created 2026-06-13T06:42:57Z; last modified 2026-09-02T02:22:41Z; 11,761 likes / 1,030,720 downloads in the 2026-10-19 API check; pipeline_tag image-text-to-text; card license other / license_name kimi-k3). GitHub MoonshotAI/Kimi-K3 has 8,945 stars (created 2026-07-27).
How big is it?
The README model summary (checked 2026-10-19) lists 2.8T total / 104B activated parameters, 896 experts with 16 selected per token, 1,048,576-token context, and native text/image (video named in the introduction). Hugging Face safetensors total 2,779,931,837,184 (~2.78T). “World’s first open 3T-class model” is a vendor sentence.
Can I use it commercially?
The Kimi K3 License is not MIT/Apache. MaaS operators over USD 20 million aggregate revenue in any consecutive 12 months need a separate Moonshot agreement before commercial use. High MAU / monthly-revenue products must display “Kimi K3”. Internal use and official/certified endpoints are carved out of those sections. Read LICENSE. Not legal advice.
Does it match ChatGPT quality?
We do not claim that. README tables versus named frontier models are vendor/self-reported. Arena.ai Code Arena WebDev (open-source filter, 2026-10-18 snapshot) lists Kimi K3 (Max) as open-source rank 2 / overall rank 15 (score 1655). That is a third-party leaderboard, not a forkalt ranking. See DeepSeek-V4.1-Flash vs Kimi K3 for sourced facts only.
Sources & verification
- https://huggingface.co/moonshotai/Kimi-K3
- https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE
- https://github.com/MoonshotAI/Kimi-K3
- https://huggingface.co/collections/moonshotai/kimi-k3
- https://platform.kimi.ai/docs/guide/kimi-k3-quickstart
- https://arena.ai/leaderboard/code/webdev?license=open-source
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.