Skip to content
Weights Not Walls
Menu

moonshotai

Kimi-K3

Reviewedaccepts images
Size
2.8T parametersfrom the GGUF metadata of a conversion, because this repo's own weights are packed: unsloth/Kimi-K3-GGUF
Licence
unknown, Kimi K3 LicenseOur reading, check the licence: Commercial use is allowed, but if you or an affiliate run a Model as a Service business and your combined revenue exceeds 20 million US dollars over any consecutive 12 months, you need a separate agreement with Moonshot AI before commercial use. Products or services with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue must prominently display "Kimi K3" in their user interface. Neither rule applies to internal use or to use through Moonshot AI's official products or certified inference partners; otherwise it is MIT style, with no duty to share changes.
On the Hub since
13 Jun 2026the date the repo was created
Context length
1M tokensfrom the model's config
Kind
Base model (no base model declared on the Hub)KimiK3ForConditionalGeneration
On Hugging Face
moonshotai/Kimi-K3734k downloads in the last 30 days, 12k likes

Scores

41 independent results on 28 eval sets
Chess PuzzlesRe-run by Epoch AI
  • effort high37.0%±4.913 Oct 2026, log
  • effort max36.0%±4.813 Oct 2026, log
  • effort low18.0%±3.913 Oct 2026, log

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_max, 79.0% (Anthropic). All 166 results

FrontierMath-ErdosRe-run by Epoch AI

Epoch's name: kimi-k3. Best closed model on this set: gpt-6.2-astra_max, 10.3% (OpenAI). All 16 results

  • effort max39.0%±7.717 Jul 2026, log

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_max, 100.0% (Anthropic). All 83 results

  • effort max72.2%±2.717 Jul 2026, Epoch AI

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_max, 95.1% (Anthropic). All 127 results

Furniture AssemblyRe-run by Epoch AI
  • effort max34.2%±6.111 Sep 2026, Epoch AI

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_max, 90.0% (Anthropic). All 41 results

GPQA diamondRe-run by Epoch AI
  • effort max93.1%±1.516 Jul 2026, log
  • effort high91.9%±1.97 Aug 2026, log
  • effort low84.8%±2.67 Aug 2026, log

Epoch's name: kimi-k3. Best closed model on this set: gpt-6.2-astra_max, 96.8% (OpenAI). All 341 results

MirrorCodeRe-run by Epoch AI
  • effort max54.4%±11.43 Sep 2026, log

Epoch's name: kimi-k3. Best closed model on this set: claude-sonnet-5-5_max, 97.8% (Anthropic). All 47 results

Mystery Game PuzzlesRe-run by Epoch AI
  • effort max26.0%±4.429 Jul 2026, Epoch AI
  • effort low12.0%±3.329 Aug 2026, log

Epoch's name: kimi-k3. Best closed model on this set: gpt-6.2-astra_max, 95.0% (OpenAI). All 148 results

OTIS Mock AIME 2024-2025Re-run by Epoch AI
  • effort max97.2%±1.116 Jul 2026, log
  • effort high93.3%±3.87 Aug 2026, log
  • effort low68.9%±7.07 Aug 2026, log

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5_high, 100.0% (Anthropic). All 318 results

SimpleQA VerifiedRe-run by Epoch AI
  • effort max50.6%±1.627 Aug 2026, log

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_max, 94.0% (Anthropic). All 105 results

APEX-AgentsReported

Epoch's name: kimi-k3. Best closed model on this set: gemini-4-carbon_unknown, 84.2% (Google DeepMind). All 66 results

ARC-AGIReported

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5_max, 98.5% (Anthropic). All 318 results

ARC-AGI-2Reported

Epoch's name: kimi-k3. Best closed model on this set: gpt-6-astra_max, 95.0% (OpenAI). All 299 results

Epoch's name: kimi-k3. Best closed model on this set: gemini-4-carbon_unknown, 58.6% (Google DeepMind). All 38 results

CritPtReported

Epoch's name: kimi-k3. Best closed model on this set: gemini-4-carbon_high, 36.0% (Google DeepMind). All 271 results

DeepSWEReported

Epoch's name: kimi-k3. Best closed model on this set: gpt-6-astra_xhigh, 74.1% (OpenAI). All 69 results

DTBenchReported

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_unknown, 99.2% (Anthropic). All 241 results

FrontierCodeReported
  • effort unknown; Harness: mini-swe-agent; Reasoning effort: none44.2%cognition.com

Epoch's name: kimi-k3. Best closed model on this set: claude-opus-5-5_medium, 54.6% (Anthropic). All 46 results

FrontierSWEReported

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_max, 69.9% (Anthropic). All 25 results

GBAEvalReported

Epoch's name: kimi-k3. Best closed model on this set: claude-opus-5-5_unknown, 90.6% (Anthropic). All 28 results

GDP.pdfReported

Epoch's name: kimi-k3. Best closed model on this set: gemini-4-carbon_high, 56.4% (Google DeepMind). All 19 results

LMCAReported

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_xhigh, 70.6% (Anthropic). All 203 results

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-1_max, 44.6% (Anthropic). All 22 results

ProofBenchReported

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-1_max, 100.0% (Anthropic). All 85 results

SimpleBenchReported

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_unknown, 94.9% (Anthropic). All 125 results

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5_high, 95.0% (Anthropic). All 29 results

WeirdMLReported

Epoch's name: kimi-k3. Best closed model on this set: gpt-6-astra_promax, 93.6% (OpenAI). All 173 results

WeirdML v3Reported

Epoch's name: kimi-k3. Best closed model on this set: claude-fable-5-2_xhigh, 65.3% (Anthropic). All 27 results

Claimed by the maker Claimed

The model card has no claimed scores in a machine-readable form (the card's model-index), so none are shown. The maker may give scores in the card text: read the card.

Other views, which we link to and do not copy: look this model up on Artificial Analysis and the Arena leaderboard.

Independent scores from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Each row links its own source. What the badges mean.

Run it yourself

file sizes read from the Hub 19 Nov 2026

Memory needed = weights + context cache + 0.50 GB for the runtime. Context cache at 8k tokens: 0.82 GB, worked out from the model's config (moonshotai/Kimi-K3), at 16-bit precision. Choose a machine above to see a verdict for each format. Method

GGUF for llama.cpp, Ollama, LM Studio and others

unsloth/Kimi-K3-GGUF (converted by unsloth), 193k downloads. Sizes are the real files on the Hub, all parts added up. For image input you also need the projector file mmproj-F16.gguf (0.84 GB), not counted below.

FormatSizeMemory neededGet it
UD-Q1_0434.3 GBfile size, 11 parts435.7 GB434.3 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-Q1_0
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-Q1_0/Kimi-K3-UD-Q1_0-*-of-00011.gguf" --local-dir Kimi-K3-GGUF
UD-TQ1_0474.0 GBfile size, 12 parts475.3 GB474.0 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-TQ1_0
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-TQ1_0/Kimi-K3-UD-TQ1_0-*-of-00012.gguf" --local-dir Kimi-K3-GGUF
UD-TQ2_0513.6 GBfile size, 13 parts514.9 GB513.6 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-TQ2_0
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-TQ2_0/Kimi-K3-UD-TQ2_0-*-of-00013.gguf" --local-dir Kimi-K3-GGUF
UD-IQ1_S553.2 GBfile size, 14 parts554.6 GB553.2 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-IQ1_S
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-IQ1_S/Kimi-K3-UD-IQ1_S-*-of-00014.gguf" --local-dir Kimi-K3-GGUF
UD-IQ1_M604.3 GBfile size, 15 parts605.6 GB604.3 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-IQ1_M
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-IQ1_M/Kimi-K3-UD-IQ1_M-*-of-00015.gguf" --local-dir Kimi-K3-GGUF
UD-IQ2_XXS662.2 GBfile size, 16 parts663.6 GB662.2 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-IQ2_XXS
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-IQ2_XXS/Kimi-K3-UD-IQ2_XXS-*-of-00016.gguf" --local-dir Kimi-K3-GGUF
UD-Q2_K_XL802.1 GBfile size, 19 parts803.4 GB802.1 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-Q2_K_XL
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-Q2_K_XL/Kimi-K3-UD-Q2_K_XL-*-of-00019.gguf" --local-dir Kimi-K3-GGUF
UD-Q4_K_XL1405.1 GBfile size, 32 parts1406.4 GB1405.1 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-Q4_K_XL
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-Q4_K_XL/Kimi-K3-UD-Q4_K_XL-*-of-00032.gguf" --local-dir Kimi-K3-GGUF
UD-Q8_K_XL1453.9 GBfile size, 34 parts1455.3 GB1453.9 + 0.82 + 0.50
Commands
llama.cpp
llama-server -hf unsloth/Kimi-K3-GGUF:UD-Q8_K_XL
Download
hf download unsloth/Kimi-K3-GGUF --include "UD-Q8_K_XL/Kimi-K3-UD-Q8_K_XL-*-of-00034.gguf" --local-dir Kimi-K3-GGUF

MLX for Apple silicon

FormatSizeMemory neededRepo
MLX mxfp4pipenetwork/Kimi-K3-REAP80-MLX-mxfp4-q8325.7 GBfile size327.0 GB325.7 + 0.82 + 0.50Download page
MLX mxfp4pipenetwork/Kimi-K3-REAP73-MLX-mxfp4-q8420.4 GBfile size421.7 GB420.4 + 0.82 + 0.50Download page

Original weights

The maker's own weights are packed into integer tensors, so the safetensors header gives no size we can trust; the repo lists the real file sizes.

hf download moonshotai/Kimi-K3 --local-dir Kimi-K3
Other conversions (16; sizes not read)

The Hub tags 57 quantised repos as conversions of this model; we read the most downloaded and show the maker's own first.

Run it by API

Hugging Face Inference Providers, fetched 19 Nov 2026, 13:58 UTC

18 providers serve this model through the Hugging Face router. Prices are US dollars per million tokens. Precision is what the provider reports; a lower precision than the original weights can change answers. Live first, then cheapest output.

ProviderInOutContextPrecisionTokens/sFirst tokenUptime 24hToolsJSON
Wafer$0.28$9.601Mmxfp4900.6 s100.0%yesyes
Relace$0.18$111Mfp4691.1 s99.9%yesno
DigitalOcean$2.55$12.91Mnot given291.7 s100.0%yesyes
Io Net$1.90$141024kfp4322.9 s99.3%yesyes
DeepInfra$2.85$14.31Mmxfp4580.9 s99.8%yesyes
Amazon Bedrock$3.00$151Mnot given952.6 s99.9%yesno
BaseTen$3.00$151Mfp8551.1 s99.6%yesno
Fireworks$3.00$151Mnot given391.6 s99.7%yesyes
InferenceNet$0.24$151Mfp4800.5 s100.0%yesyes
Modal$3.00$151Mmxfp4670.8 s100.0%yesyes
Moonshot AIthe maker$3.00$151Mnot given363.4 s99.4%yesyes
Nebius$3.00$151Mfp41151.8 s100.0%yesyes
Parasail$3.00$151Mfp4261.2 s100.0%yesyes
River$3.00$151Mnot given1240.3 s99.6%noyes
Together$3.00$151Mnot given621.1 s100.0%yesyes
Sail Research$0.34$16.61Mfp4580.7 s100.0%yesyes
Makora$1.00$181Mfp4361.4 s99.9%yesyes
featherless-ai???not given???not givennot given

Through the router, with a Hugging Face token (the router bills your Hugging Face account). Change the name after the colon to pick another provider.

curl https://router.huggingface.co/v1/chat/completions \ -H "Authorization: Bearer $HF_TOKEN" -H "Content-Type: application/json" \ -d '{"model": "moonshotai/Kimi-K3:wafer", "messages": [{"role": "user", "content": "Hello"}]}'

Not built yet: one key for every open model, with BlockRun, and our own tested-quality score per provider. Until then, precision and measured speed are the best guide to a provider here.

Fine-tunes of this model

from our snapshot of the Hub

Repos that declare this model as their base, most downloaded first. Listing is not a recommendation; unreviewed ones have not been checked.

Planned for this page Not built yet

Try it in private chat
A link to Solana AI's confidential chat for the models it serves. This model is not listed as served there.
Fine-tune it
Fine-tuning a few base models from this page and publishing the result here.
Creators and earnings
Pages for people who publish fine-tunes and adapters of this model, and what they earn.
Fund and bounties
Bounties for new open models, paid from a public fund. How the fund will work.
Contribute compute
How to give compute to open model work, with Akash.

Model data from the Hugging Face Hub 19 Nov 2026, 13:43 UTC. The repo. Found a mistake on this page? Tell us.