Models › Kimi-K3 family

moonshotai Kimi K3 (Modified MIT) (kimi-k3) out 4 months ago

2.78T parameters, images and text in, text out, 1M tokens of context. 22 providers serve it from $0.24 per million input tokens.

Verified 2026-10-25 Rank #7 in LLMs · score 131.7 method v1 Sources 5 Downloads 30d 909.8k · #119 of 619 tracked

Changed this week: Cheapest input price fell 57.8%: $0.536 → $0.226 per 1M (OpenRouter).2026-10-24 · Cheapest input price fell 55.6%: $0.536 → $0.238 per 1M (OpenRouter).2026-10-24 · 2 new quantized builds on the Hub (58 in all).2026-10-24 · 1 more

Can I run it?1.67TB at INT4Estimateneeds a multi-GPU nodemethodCommunity: Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs →
What does it cost?$0.24/1M incheapest of 20Wafermxfp42026-10-24
Can I use it?Kimi K3 (Modified MIT) (kimi-k3)MIT plus a 'Kimi K3' attribution requirement for products above 100M MAU or US$20M monthly revenue (license text in repo)licence file2026-10-24
Does it work?5/5probes · 59 tok/sfirst lookparasail2026-10-07
Do people use it?909.8kdownloads, 30 days#119 of 619 we track · top 19%+68 likes, downloads −13% this week11,818 likesHugging Face2026-10-24

The ringer: four playful tests

A pelican steals a bicycle from a supermarket

low reasoning✕ no SVG

No drawable SVG in the reply.

Draw a pelican stealing a bicycle from a supermarket. Reply with one complete SVG document only, no prose and no code fences. Use viewBox="0 0 200 150" and a transparent background. Use only these colours for every fill and stroke: #06e0c0 #ffc298 #cd85f9 #f1f2f2.

Finish the dad joke

as received● 1 line · 6 words

Why did the open-weights model cross the road?

To get to the open source.

Finish this dad joke. Reply with the punchline only, on one line, without quotes: Why did the open-weights model cross the road?

A parasailer, in ASCII

as received● 12×24 fits 12×32
        .-~~~~-.
     .-'        '-.
    /  \      /    \
   |    \    /      |
    \    \  /      /
     '-.  \/   .-'
        | /\  |
        |/  \ |
        | o  |
~~~~~~~~|/|\ |~~~~~~~~
 ~  ~  ~ / \ ~  ~  ~
~~~~~~~~~~~~~~~~~~~~~~~~
Draw ASCII art of a parasailer over the sea. At most 12 lines and 32 columns. Reply with the art only: no code fences, no explanation.

Animate a bouncing ball

low reasoning● 8 frames · 16×16
Make a 16x16 pixel sprite animation of a ball bouncing, 8 frames. Reply with JSON only, no code fences: {"palette": {"a": "#rrggbb", ...}, "fps": <number>, "frames": [[16 strings of 16 characters], ...]}. At most 4 palette colours with single-letter keys; "." means transparent; exactly 8 frames.
1 of 4 · swipe for more

The ringer: the same four playful tests for every model, replies shown as received and checked by code. A joke, not a benchmark; never ranked.3 of 4 pass · parasail-kimi-k3 on parasail · run 2026-10-09 · temperature 0 · cost $0.121 · 2026-10 thing: supermarket · all models →

Worth reading

Picked by rule from 35 tracked pieces: the lab's own post, the most shared, a field report, the latest.

The lab's own post

Kimi K3: Open Frontier Intelligence

Kimi K3 is a 2.8T-parameter open model featuring native vision and a 1-million-token context window.

kimi.com 2026-07-16 2,107 HN pts · 1,214 comments 2.6k words
Latest · 47 days ago

Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs

A fork of Deltafin runs the full 2.8-trillion-parameter Kimi K3 model on a single M5 Max MacBook Pro with 128 GB RAM, streaming 1.45 TB of expert weights from four SSDs.

  • The slowest read among a layer's 16 experts sets the pace, not total bandwidth.
  • Streaming setup initially runs far more slowly when routes have no local cache yet.
github.com 2026-09-08 279 HN pts · 156 comments 2.1k words

25 best of 35 tracked

The 4 picks above, plus these 21. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleKimi K3-256kkimi.com · 2026-07-29 · 492 HN pts · official
  2. ArticleKimi K3 Is Competitive with Fable; Kimi K3 and Fable Is SoTAfireworks.ai · 2026-07-21 · 877 HN pts
  3. ArticleKimi K3 Architecture Overview and Notessebastianraschka.com · 2026-07-28 · 507 HN pts
  4. CodeKimi-K3 Technical Report [pdf]github.com · 2026-07-27 · 391 HN pts
  5. ArticleThe Kimi K3 Momentstephen.bochinski.dev · 2026-07-18 · 636 HN pts
  6. ArticleKimi K3, Qwen 3.8, and Anthropic's (Potential) Unravellingemergingtrajectories.com · 2026-07-20 · 371 HN pts
  7. CodeRunning Kimi K3 on a M1 Maxgithub.com · 2026-07-28
  8. ArticleRunning Kimi K3 on MI355X at Better Performance per Dollar Than B300wafer.ai · 2026-08-02
  9. DiscussionKimi K3 exploited the latest Redis servertwitter.com · 2026-07-23
  10. DiscussionMoonshot AI suspends new subscriptions due to Kimi K3 demandtwitter.com · 2026-07-19
  11. ArticleKimi K3 Now Available via Telnyx Inference APItelnyx.com · 2026-07-27
  12. ArticleSelf-hosting Kimi K3: 20% more hardware cost, 20% better task resolutionaistack.imec-int.com · 2026-07-29
  13. ArticleKimi K3: second only to Fable 5 on AA-Briefcaseartificialanalysis.ai · 2026-07-22
  14. ArticleUK AISI / Caisi Preliminary Assessment of Kimi K3's Cyber Capabilitiesnist.gov · 2026-07-25
  15. ArticleKimi K3 built a Windows XP in browserwindows-xp.kimi.site · 2026-07-25
  16. ArticleKimi K3 Intelligence, Performance and Price Analysisartificialanalysis.ai · 2026-07-16
  17. ArticleKimi K3: The open-weights escalationinterconnects.ai · 2026-07-20
  18. DiscussionKimi K3 is now #1 in the Front end Code Arena with 1679 pts, surpassing Fable 5twitter.com · 2026-07-16
  19. ArticleKimi K3's Design Secret May Be in Its Thinking Tracesnotes.designarena.ai · 2026-07-30
  20. ArticleLLM City – 3D render of all Kimi K3's weights as 2.5mm tilesmagik.net · 2026-08-17
  21. ArticleKimi K3 is not cheapalexinch.com · 2026-07-26

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

Prices for the same weights vary 14×

20 providers with a live price, from $0.24 to $3.45 per million input tokens; the median is $2.42. One dot per provider, cheapest in teal.

$0.010$0.10$1.00$10.00Wafer: $0.24 in / $16.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3InferenceNet: $0.24 in / $10.80 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Makora: $0.65 in / $15.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Sail Research: $1.20 in / $5.50 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3AkashML: $1.20 in / $14.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Morph: $1.42 in / $14.90 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Decart: $1.65 in / $8.25 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Chutes: $1.85 in / $9.50 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Fireworks AI: $2.10 in / $10.50 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/moonshotai/Kimi-K3?expand[]=inferenceProviderMappingDeepInfra: $2.28 in / $11.40 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/moonshotai/Kimi-K3?expand[]=inferenceProviderMappingPhala: $2.55 in / $12.75 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3DigitalOcean: $2.55 in / $12.95 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Together AI: $3.00 in / $15.00 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/moonshotai/Kimi-K3?expand[]=inferenceProviderMappingBaseten: $3.00 in / $15.00 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/moonshotai/Kimi-K3?expand[]=inferenceProviderMappingParasail: $3.00 in / $15.00 out per 1M · observed 2026-10-24 · https://docs.parasail.io/parasail-docs/billing/pricingParetoAlpha: $3.00 in / $15.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Amazon Bedrock: $3.00 in / $15.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Modal: $3.00 in / $15.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3+1Alibaba: $3.45 in / $17.25 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3cheapest $0.24median $2.42
$0.010$0.10$1.00$10.00Wafer: $0.24 in / $16.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3InferenceNet: $0.24 in / $10.80 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Makora: $0.65 in / $15.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Sail Research: $1.20 in / $5.50 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3AkashML: $1.20 in / $14.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Morph: $1.42 in / $14.90 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Decart: $1.65 in / $8.25 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Chutes: $1.85 in / $9.50 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Fireworks AI: $2.10 in / $10.50 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/moonshotai/Kimi-K3?expand[]=inferenceProviderMappingDeepInfra: $2.28 in / $11.40 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/moonshotai/Kimi-K3?expand[]=inferenceProviderMappingPhala: $2.55 in / $12.75 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3DigitalOcean: $2.55 in / $12.95 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Together AI: $3.00 in / $15.00 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/moonshotai/Kimi-K3?expand[]=inferenceProviderMappingBaseten: $3.00 in / $15.00 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/moonshotai/Kimi-K3?expand[]=inferenceProviderMappingParasail: $3.00 in / $15.00 out per 1M · observed 2026-10-24 · https://docs.parasail.io/parasail-docs/billing/pricingParetoAlpha: $3.00 in / $15.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Amazon Bedrock: $3.00 in / $15.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3Modal: $3.00 in / $15.00 out per 1M · observed 2026-10-24 · https://openrouter.ai/moonshotai/kimi-k3+2cheapest $0.24 · median $2.42 (white tick)
#Cheapest 5In /1MOut /1MContextQuantSeen
1 Wafer $0.24 $16.00 1M mxfp4 Oct 24
2 InferenceNet $0.24 $10.80 1M fp4 Oct 24
3 Makora $0.65 $15.00 1M fp4 Oct 24
4 Sail Research $1.20 $5.50 1M fp4 Oct 24
5 AkashML $1.20 $14.00 1M fp4 Oct 24

All 22 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Widely served (22 API providers); heavily run locally (58 quantized builds); a big base for builders (54 fine-tunes, adapters and merges).

  • As a service22AkashML, Alibaba, Amazon Bedrock, Baseten +18
  • On own hardware58quantized builds
  • Built on5452 fine-tunes · 2 adapters · 0 merges
  • In apps86Hugging Face Spaces

Top Spaces running it: embedl/hfviewer ♥ 181 · akhaliq/Kimi-K3 ♥ 19 · discord-community/HuggingMod ♥ 17 · HapppyHooochie/Kimi-K3-Abliterated-Demo ♥ 10

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Too big for one 8×80 GB node at any precision

About 1.67 TB at INT4, 3.34 TB at FP8, 6.67 TB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.67 TB of weights (estimate; KV cache and activations extra)INT41.67 TBFP8: about 3.34 TB of weights (estimate; KV cache and activations extra)FP83.34 TBBF16: about 6.67 TB of weights (estimate; KV cache and activations extra)BF166.67 TB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.67 TB of weights (estimate; KV cache and activations extra)INT41.67 TBFP8: about 3.34 TB of weights (estimate; KV cache and activations extra)FP83.34 TBBF16: about 6.67 TB of weights (estimate; KV cache and activations extra)BF166.67 TB

Estimate params × bytes per param × 1.2 · how we estimate

Day-one receipts

The probe replies as received on 2026-10-07 from parasail, graded by code. Receipts, not a score.

Write a function● pass
def is_balanced(s: str) -> bool:
    pairs = {')': '(', ']': '[', '}': '{'}
    stack = []
    for ch in s:
        if ch in '([{':
            stack.append(ch)
        elif ch in ')]}':
            if not stack or stack.pop() != pairs[ch]:
                return False
    return not stack
Bracket-balance function passes 9 hidden asserts4.1 s
Tool call● pass
{"city": "Lisbon", "unit": "c"}
Emits a correct function call (get_weather, city=Lisbon)1.8 s
Strict JSON● pass
{"name":"Foo-9B-Instruct","license":"Apache-2.0","parameters_b":9.2,"context_tokens":131072,"languages":["English","French","German","Japanese"]}
Structured extraction to strict JSON (5 fields)2.5 s
Speed run● pass
{
  "providers": [
    {
      "provider": "parasail",
      "endpoint_model": "parasail-kimi-k3",
      "source_url": "https://api.parasail.io/v1/chat/completions",
      "ttft_ms_median": 13942,
      "ttft_any_ms_median": 589,
      "tokens_per_s_median": 88.3,
      "runs": [
        {
          "ttft_s": 13.942,
          "ttft_any_s": 0.994,
          "tokens_per_s": 103.8,
          "latency_s": 14.44,
          "tokens_out": 1500,
          "reasoning_chars": 5407,
          "finish": "length"
        },
        {
          "ttft_s": 9.107,
          "ttft_any_s": 0.589,
          "tokens_per_s": 88.3,
          "latency_s": 10.09,
          "tokens_out": 891,
          "reasoning_chars": 3398,
          "finish": "stop"
        },
        {
          "ttft_s": 18.848,
          "ttft_any_s": 0.557,
          "tokens_per_s": 70.8,
          "latency_s": 20.28,
          "tokens_out": 1435,
          "reasoning_chars": 4930,
          "finish": "stop"
        }
      ]
    }
  ],
  "note": "Single client, wall clock; tok/s counts every output token including hidden reasoning."
}
Three-tool errand● pass
{
  "calls": [
    {
      "name": "get_weather",
      "args": {
        "city": "Lisbon",
        "date": "2026-10-14"
      }
    },
    {
      "name": "check_calendar",
      "args": {
        "date": "2026-10-14"
      }
    },
    {
      "name": "book_table",
      "args": {
        "date": "2026-10-14",
        "party_size": 2,
        "restaurant": "Cervejaria Ramiro",
        "time": "19:30"
      }
    }
  ],
  "checks": {
    "get_weather": true,
    "check_calendar": true,
    "book_table": true
  },
  "final": "All done! Here's a summary:\n\n- **Weather in Lisbon on 2026-10-14:** Clear, high of 22°C, only 5% chance of rain — no rain expected ☀️\n- **Free evening slot found:** 19:30–21:30\n- **Reservation:** Table for 2 at Cervejaria Ramiro, 19:30 on 2026-10-14\n\n**Confirmation code: TBL-4471**\n\nEnjoy your evening — the seafood there is fantastic!",
  "fixture": "receipts.errand.v1"
}
calls in order · args 3/3 · confirmed · 2026-10-095.2 s
Needle heatmap● pass
{
  "lengths_k": [
    8,
    32,
    64
  ],
  "depths": [
    0.1,
    0.5,
    0.9
  ],
  "cells": [
    [
      1,
      1,
      1
    ],
    [
      1,
      1,
      1
    ],
    [
      1,
      1,
      1
    ]
  ],
  "detail": [
    {
      "k": 8,
      "depth": 0.1,
      "recalled": true,
      "reply": "amber-falcon-7731",
      "latency_s": 1.91,
      "finish": "stop"
    },
    {
      "k": 8,
      "depth": 0.5,
      "recalled": true,
      "reply": "cobalt-marten-7828",
      "latency_s": 1.53,
      "finish": "stop"
    },
    {
      "k": 8,
      "depth": 0.9,
      "recalled": true,
      "reply": "saffron-newt-7925",
      "latency_s": 1.32,
      "finish": "stop"
    },
    {
      "k": 32,
      "depth": 0.1,
      "recalled": true,
      "reply": "jade-lynx-8022",
      "latency_s": 3.85,
      "finish": "stop"
    },
    {
      "k": 32,
      "depth": 0.5,
      "recalled": true,
      "reply": "umber-kestrel-8119",
      "latency_s": 3.7,
      "finish": "stop"
    },
    {
      "k": 32,
      "depth": 0.9,
      "recalled": true,
      "reply": "coral-heron-8216",
      "latency_s": 2.21,
      "finish": "stop"
    },
    {
      "k": 64,
      "depth": 0.1,
      "recalled": true,
      "reply": "slate-badger-8313",
      "latency_s": 5.44,
      "finish": "stop"
    },
    {
      "k": 64,
      "depth": 0.5,
      "recalled": true,
      "reply": "ochre-otter-8410",
      "latency_s": 4.15,
      "finish": "stop"
    },
    {
      "k": 64,
      "depth": 0.9,
      "recalled": true,
      "reply": "teal-ibis-8507",
      "latency_s": 3.67,
      "finish": "stop"
    }
  ]
}

Every probe, with timings and cost →

The story so far

Weights, every dated article, and repo commits on one rail.

10 dated articles and 8 repo commits so far

Teal is the weights (2026-06-13), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

What people built on it

112 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

112 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

58 quantized (Hugging Face model tree)52 fine-tunes (Hugging Face model tree)2 adapters (Hugging Face model tree)58 quantized52 fine-tunes2 adapters
58 quantized (Hugging Face model tree)52 fine-tunes (Hugging Face model tree)2 adapters (Hugging Face model tree)58 quantized52 fine-tunes2 adapters

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

11 rows

Benchmarks 11 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (11)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported AA-LCR · Read long documents74.7—model author2026-10-07
author-reported CritPt23.4—model author2026-10-07
author-reported DeepSWE · Write code with agents67.5—model author2026-10-07
author-reported GPQA Diamond · Reason through hard problems93.5—model author2026-10-07
author-reported HLE-Full43.5 / 56.0—model author2026-10-07
author-reported ProgramBench77.8—model author2026-10-07
author-reported apex-agents (mercor/apex-agents)41—Model Card2026-10-24
author-reported deep_swe (datacurve/deep-swe)67.5—Model Card2026-10-24
author-reported diamond (Idavidrein/gpqa)93.5—Model Card2026-10-24
author-reported hle (cais/hle)56—Model Card2026-10-24
author-reported toolathlon_verified (hkust-nlp/Toolathlon)76.5—moonshotai/Kimi-K3 model card2026-08-20

What the card says

1 cited sentence

Summary generated, cited

The architecture utilizes a Mixture-of-Experts framework and supports a 1-million-token context window.[1]

  1. “1-million-token context window”

Generated 2026-10-07 14:59 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

7 use cases

Facts

21 fields

Listed in LLMs

Facts

Pipeline tagimage-text-to-text
Librarytransformers
ArchitecturesKimiK3ForConditionalGeneration
Model typekimi_k3
Dtypebfloat16
Params by dtypeBF16: 57,179,884,544 · F32: 11,122,432 · U8: 2,722,740,830,208
Size bandXXL
MoEyes
Experts896
Context length1,048,576 (config.json max_position_embeddings)
LicenseKimi K3 (Modified MIT) (kimi-k3) (other)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · KimiK3ForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · KimiK3ForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Kimi-K3: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)6.67 TB · Cluster (≥640 GB)
FP8 (1 byte/param)3.34 TB · Cluster (≥640 GB)
INT4 (0.5 bytes/param)1.67 TB · Cluster (≥640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

First look, every probe

6 probes

First look measured

A fixed probe set run against a live API endpoint: the same questions for every model, answers graded by code, nothing judged by a model. Latencies are single-request wall clock from one client location, not a benchmark.

via parasail parasail-kimi-k3 5/5 probes pass first visible token 25.6 s end-to-end 59 tok/s (median of 3)
ProbeResultLatencyReply (truncated)
Self-description info 2.6 s I'm Kimi, developed by Moonshot AI (月之暗面). I don't know my training data cutoff.
Tool call pass 1.8 s {"city": "Lisbon", "unit": "c"}
Strict JSON pass 2.5 s {"name":"Foo-9B-Instruct","license":"Apache-2.0","parameters_b":9.2,"context_tokens":131072,"languages":["English","French","German","Japanese"]}
Code (hidden tests) pass 4.1 s def is_balanced(s: str) -> bool: pairs = {')': '(', ']': '[', '}': '{'} stack = [] for ch in s: if ch in '([{': stack.append(ch) elif ch in ')]}': if not stack or stack.pop() != pairs[ch]: return False return not stack
Multi-step arithmetic pass 1.9 s 16
Needle recall at 64k tokens pass 6.7 s amber-falcon-7731

Probe set firstlook.v1 · observed 2026-10-07 20:12 UTC · endpoint · provider docs · cost of this run $0.23

All variants

40 variants

Lineage

Original release (shipped quantized by the lab) Family: moonshotai/Kimi-K3 112 derivatives on the Hub

quantized: 58finetune: 52adapter: 2

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
unsloth/Kimi-K3-GGUF quantized GGUF — — other unsloth 402 1.3M
inference-optimization/Kimi-K3-0.40B finetune — 396M S mit inference-optimization 83 88.5k
Inferact/Kimi-K3-DSpark finetune — 3.56B M other Inferact 55 75.8k
nvidia/Kimi-K3-NVFP4 quantized NVFP4 1.42T XXL other nvidia 42 46.1k
pipenetwork/Kimi-K3-REAP80-MLX-mxfp4-q8 quantized MLX 601B XXL other pipenetwork 16 31.5k
AtomicChat/Kimi-K3-GGUF quantized GGUF — — other AtomicChat 17 19.9k
RedHatAI/Kimi-K3-NVFP4 quantized NVFP4 1.59T XXL other RedHatAI 13 20k
pipenetwork/Kimi-K3-REAP73-MLX-mxfp4-q8 quantized MLX 792B XXL other pipenetwork 9 19k
modal-labs/Kimi-K3-DFlash finetune — 2.60B M other modal-labs 4 19.1k
mgoin/Kimi-K3-pruned75 finetune — 737B XXL other mgoin 1 17.1k
unsloth/Kimi-K3 quantized — 2.78T XXL other unsloth 257 2.9k
Ryanchen911/Kimi-K3-Uncensored-GGUF quantized GGUF — — other Ryanchen911 55 11k
Uniboshi/Kimi-K3-Abliterated-V1 finetune — 2.78T XXL other Uniboshi 146 5.2k
pipenetwork/Kimi-K3-REAP73-zh-code-MLX-mxfp4-q8 quantized MLX 792B XXL other pipenetwork 8 11.3k
kernelpool/Kimi-K3-2bit-UVMAX quantized — 2.78T XXL other kernelpool 5 9.8k
avlp12/Kimi-K3-Alis-MLX-Dynamic-2.10bpw quantized MLX 449B XXL other avlp12 4 8.7k
GrEarl/Kimi-K3-GGUF quantized GGUF — — other GrEarl 30 5.6k
pipenetwork/Kimi-K3-REAPgraded-MLX-mxfp4-q8 quantized MLX 1.05T XXL other pipenetwork 1 6.8k
mlx-community/Kimi-K3-mlx-reap160-2bit finetune mlx 543B XXL other mlx-community 2 6.3k
Kuberwastaken/Kimi-K3-GGUF quantized GGUF — — other Kuberwastaken 14 5.6k
bullerwins/Kimi-K3-GGUF quantized GGUF — — — bullerwins 0 5.7k
RedHatAI/Kimi-K3-speculator.dspark finetune — 4.74B M apache-2.0 RedHatAI 8 5k
GrEarl/Kimi-K3-GGUF-IQ1_S quantized GGUF — — other GrEarl 22 3.8k
RedHatAI/Kimi-K3-FP8-BLOCK quantized -FP8 2.78T XXL other RedHatAI 12 3.8k
catalystsec/Kimi-K3-Audit-451E-MLX-MXFP4 quantized MLX 1.43T XXL other catalystsec 1 3.6k
catalystsec/Kimi-K3-Audit-240E-MLX-MXFP4 quantized MLX 786B XXL other catalystsec 0 3.6k
qtum/Kimi-K3-GGUF quantized GGUF — — other qtum 0 3.4k
nota-ai/Kimi-K3-Nota-Global-Pruned-50 finetune — 1.42T XXL other nota-ai 15 2.4k
6block/Kimi-K3-GGUF quantized GGUF — — other 6block 1 2.9k
vcruz305/Kimi-K3-GGUF quantized GGUF — — other vcruz305 6 2.4k
prometheusAIR/Kimi-K3-REAP55-GGUF quantized GGUF — — other prometheusAIR 15 1.4k
lightseekorg/kimi-k3-dspark finetune — 3.56B M other lightseekorg 2 1.8k
abacusai/Smaug-Agentic finetune — 2.78T XXL other abacusai 25 669
vessl/Kimi-K3-W4AFP8 quantized — 1.42T XXL other vessl 12 1.3k
riverclouds/Kimi-K3-8L-dummy finetune — — — other riverclouds 0 1.8k
Inferact/Kimi-K3-DSpark-Block5 finetune — 2.85B M — Inferact 2 1.6k
lightseekorg/kimi-k3-eagle3-mla finetune — 3.02B M other lightseekorg 1 1.6k
DevQuasar/moonshotai.Kimi-K3-GGUF quantized GGUF — — — DevQuasar 6 1.4k
PatronusAI/kimi-k3-nvfp4 quantized nvfp4 1.42T XXL — PatronusAI 2 1.2k
SHSLab/Kimi-K3-Abliterated finetune — 2.78T XXL other SHSLab 12 705

Repository changelog

8 commits
changed in the last 7 daysNew providers: 2026-10-23Price moves: 2026-10-24, 2026-10-24Other: 2026-10-24all changes

Repository changelog 8 commits

DateCommitSHA
2026-09-02 Add Toolathlon-Verified evaluation result (#183) f831ab6
2026-08-20 Update encoding_k3.py (#173) a590ce0
2026-07-27 Add community evaluation results for GPQA, HLE, DEEP-SWE, APEX-AGENTS (#37) 9f62e4e
2026-07-27 Update README.md (#33) 301be1b
2026-07-27 [HF Staff] this tag is required to enable inference providers, please merge! (#30) a7b6c02
2026-07-27 Update README.md 2496450
2026-07-27 Update README.md dc514c7
2026-07-27 Initial commit c5d1dd4

All providers

22 providers · 22 endpoints

Where to run 22 providers · 22 endpoints

Hosted endpoints observed for Kimi-K3, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.23/M.

ProviderModel refContextQuant$/M in$/M outtok/sTTFTToolsStatusObserved
Relace
openrouter
moonshotai/kimi-k3 1M fp4 $0.23 $7.70 — — yes status -2 2026-10-24
Wafer
openrouter
moonshotai/kimi-k3 1M mxfp4 $0.24 $16.00 — — yes live 2026-10-24
InferenceNet
openrouter
moonshotai/kimi-k3 1M fp4 $0.24 $10.80 — — yes live 2026-10-24
Makora
openrouter
moonshotai/kimi-k3 1M fp4 $0.65 $15.00 — — yes live 2026-10-24
Sail Research
openrouter
moonshotai/kimi-k3 1M fp4 $1.20 $5.50 — — yes live 2026-10-24
AkashML
openrouter
moonshotai/kimi-k3 1M fp4 $1.20 $14.00 — — yes live 2026-10-24
Morph
openrouter
moonshotai/kimi-k3 1M fp8 $1.42 $14.90 — — yes live 2026-10-24
Decart
openrouter
moonshotai/kimi-k3 1M mxfp4 $1.65 $8.25 — — yes live 2026-10-24
Chutes
openrouter
moonshotai/kimi-k3 1M mxfp4 $1.85 $9.50 — — — live 2026-10-24
Fireworks AI
hf, openrouter
accounts/fireworks/models/kimi-k3 1M unknown $2.10 $10.50 30 1066 ms yes live 2026-10-25
DeepInfra
hf, deepinfra, openrouter
moonshotai/Kimi-K3 1M mxfp4 $2.28 $11.40 62 451 ms yes live 2026-10-25
Phala
openrouter
moonshotai/kimi-k3 1M unknown $2.55 $12.75 — — yes live 2026-10-24
DigitalOcean
openrouter
moonshotai/kimi-k3 1M unknown $2.55 $12.95 — — yes live 2026-10-24
Together AI
hf, openrouter
moonshotai/Kimi-K3 1M unknown $3.00 $15.00 36.4 495 ms yes live 2026-10-25
Baseten
hf, openrouter
moonshotai/Kimi-K3 1M fp8 $3.00 $15.00 43 1824 ms yes live 2026-10-25
Parasail
parasail, openrouter
parasail-kimi-k3 1M fp4 $3.00 $15.00 — — yes serverless 2026-10-24
ParetoAlpha
openrouter
moonshotai/kimi-k3 1M unknown $3.00 $15.00 — — yes live 2026-10-24
Amazon Bedrock
openrouter
moonshotai/kimi-k3 1M unknown $3.00 $15.00 — — yes live 2026-10-24
Modal
openrouter
moonshotai/kimi-k3 1M mxfp4 $3.00 $15.00 — — yes live 2026-10-24
Moonshot AI
openrouter
moonshotai/kimi-k3 1M mxfp4 $3.00 $15.00 — — yes live 2026-10-24
Alibaba
openrouter
moonshotai/kimi-k3 1M unknown $3.45 $17.25 — — yes live 2026-10-24
Featherless
hf
moonshotai/Kimi-K3 — — — — — — — live 2026-10-25

Articles and coverage

35 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official2

kimi.com2026-07-16officialauthorityHN front pageHN 2,107 pts · 1,214 comments2,558 words
Kimi K3 is a 2.8T-parameter open model featuring native vision and a 1-million-token context window. It utilizes Kimi Delta Attention and Attention Residuals alongside Stable LatentMoE, activating 16 out of 896 experts to achieve a 2.5x improvement in scaling efficiency. The model demonstrates frontier-level performance in long-horizon coding, chip design, and scientific research.written, reviewed, cited
kimi.com2026-07-29officialauthorityHN front pageHN 492 pts · 157 comments1,561 words
K3.1 Pro is now fully rolled out in Kimi Code as the first model of the new K3.1 series with a 1M context window. K3.1 Pro Highspeed delivers approximately 2–3× the speed of K3.1 Pro while maintaining identical model capabilities.generated, cited

Analysis8

github.com2026-09-08authorityHN front pageHN 279 pts · 156 comments2,117 words
A fork of Deltafin runs the full 2.8-trillion-parameter Kimi K3 model on a single M5 Max MacBook Pro with 128 GB RAM, streaming 1.45 TB of expert weights from four SSDs. One cold run with the exact prompt yields 1.00 tokens per second for 512-token answers and 0.96 tok/s on a public 17-token prompt, with per-run logs available in k3-public-bench/. A 512-token prompt takes approximately 6.3 minutes to generate the first token because prefill re-reads each layer's experts 8×.written, reviewed, cited
  • The slowest read among a layer's 16 experts sets the pace, not total bandwidth.
  • Streaming setup initially runs far more slowly when routes have no local cache yet.
  • The native server refuses anything it cannot honor exactly with a normal OpenAI-shaped error.
fireworks.ai2026-07-21authorityHN front pageHN 877 pts · 450 comments895 words
Fireworks benchmarks Kimi K3 against Fable 5 across 1,030 agentic tasks, finding that routing between the two achieves 93% accuracy. This approach is up to 50 times more cost effective than using Fable alone for long loops, with K3 handling symbolic math and Fable excelling in multi-language coding. An oracle router selects K3 for 72-96% of task traffic, leveraging its efficiency of roughly 55 turns and 1.3M tokens per SWE task versus Fable's 21 turns and 130K.written, reviewed, cited
  • K3 clears specific tasks Fable fails, including a 7z hash and FEAL cryptanalysis.
  • Tasks with extra turns generally mean more wall-clock time per run i.e. slower runs.
sebastianraschka.com2026-07-28authorityHN front pageHN 507 pts · 111 comments398 words
Kimi K3 is a 2.8T parameter open-weight model that replaces standard components with efficiency-tweaked versions like LatentMoE and multi-head latent attention. It introduces attention residuals to improve the residual path, adding about 4% in training cost and 2% in inference cost. The model removes all RoPE layers and uses NoPE everywhere, marking it as the first frontier-level architecture to do so.written, reviewed, cited
simonwillison.net2026-07-17authorityHN front pageHN 407 pts · 223 comments1,140 words
Moonshot AI announced Kimi K3, a 2.8 trillion parameter model available via API with open weights promised by July 27, 2026. The model costs $3/million input tokens and $15/million output tokens, consuming 13,241 reasoning tokens to generate a simple SVG for a total cost of 25 cents. While it surpasses Claude Opus 4.8 and GPT-5.5 high on benchmarks, it loses to Claude Fable 5 and GPT-5.6 Sol.written, reviewed, cited
  • Kimi K3's tokenizer counts the 'hi' prompt as 86 tokens, suggesting an 85 token hidden system prompt.
  • The model is the most expensive released by a Chinese AI lab to date at $3/million input tokens.
  • K3 currently only has one thinking effort level called 'max' rather than multiple options.
github.com2026-07-31authorityHN front pageHN 341 pts · 166 comments5,646 words
github.com2026-07-27authorityHN front pageHN 391 pts · 183 comments
stephen.bochinski.dev2026-07-18HN front pageHN 636 pts · 613 comments536 words
emergingtrajectories.com2026-07-20HN front pageHN 371 pts · 336 comments1,244 words
All coverage tracked for this model (35) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDsgithub.com2026-09-082026-10-07hn
LLM City – 3D render of all Kimi K3's weights as 2.5mm tilesmagik.net2026-08-172026-10-07hn
Kimi K3 review: architecture, benchmarks and open weightsneurohive.io2026-08-032026-10-07exa
Beyond Bigger MoE: How Kimi K3 Scales Context, Depth, and Agents | Andrey Lukyanenkoandlukyane.com2026-08-032026-10-07exa
Running Kimi K3 on MI355X at Better Performance per Dollar Than B300wafer.ai2026-08-022026-10-07hn
Run Kimi K3 using 29 GB of RAM at 0.50 tok/sgithub.com2026-07-312026-10-07hn
Kimi K3's Design Secret May Be in Its Thinking Tracesnotes.designarena.ai2026-07-302026-10-07hn
How Kimi K3 Engineered Its Way to the Frontiercodepointer.substack.com2026-07-302026-10-07exa
Inside Kimi K3 Technical Report | Towards AItowardsai.com2026-07-302026-10-07exa
Kimi K3-256kkimi.com official2026-07-292026-10-07hn
Self-hosting Kimi K3: 20% more hardware cost, 20% better task resolutionaistack.imec-int.com2026-07-292026-10-07hn
Running Kimi K3 on a M1 Maxgithub.com2026-07-282026-10-07hn
Kimi K3 Architecture Overview and Notessebastianraschka.com2026-07-282026-10-07hn
Kimi K3 Now Available via Telnyx Inference APItelnyx.com2026-07-272026-10-07hn
Kimi-K3 Technical Report [pdf]github.com2026-07-272026-10-07hn
Kimi K3, the Full Review: The Weights Are Out. Here's What Moonshot Didn't Want Graded. | Omniscient Mediaomniscient.media2026-07-272026-10-07exa
Kimi K3 is not cheapalexinch.com2026-07-262026-10-07hn
Kimi K3 after a week of real project use: what works, what doesn'tchenchen.guru2026-07-262026-10-07hn
Kimi K3 built a Windows XP in browserwindows-xp.kimi.site2026-07-252026-10-07hn
UK AISI / Caisi Preliminary Assessment of Kimi K3's Cyber Capabilitiesnist.gov2026-07-252026-10-07hn
Kimi K3 exploited the latest Redis servertwitter.com2026-07-232026-10-07hn
Kimi K3: second only to Fable 5 on AA-Briefcaseartificialanalysis.ai2026-07-222026-10-07hn
Kimi K3 deep dive: how far the 2.8T open flagship sits from Claude and GPT — AI Hippoai-hippo.com2026-07-222026-10-07exa
Kimi K3 Is Competitive with Fable; Kimi K3 and Fable Is SoTAfireworks.ai2026-07-212026-10-07hn
Kimi K3 just fixed 15 critical security bugs that Codex and Fable refusedtwitter.com2026-07-202026-10-07hn
Kimi K3: The open-weights escalationinterconnects.ai2026-07-202026-10-07hn
Kimi K3, Qwen 3.8, and Anthropic's (Potential) Unravellingemergingtrajectories.com2026-07-202026-10-07hn
Tested Kimi K3 for Codingdotnetperls.com2026-07-202026-10-07hn
Moonshot AI suspends new subscriptions due to Kimi K3 demandtwitter.com2026-07-192026-10-07hn
The Kimi K3 Momentstephen.bochinski.dev2026-07-182026-10-07hn
Kimi K3: Open Frontier Intelligence at 2.8 Trillion Parameters | Gradients & Tidesjianyuh.github.io2026-07-182026-10-07exa
Kimi K3, and what we can still learn from the pelican benchmarksimonwillison.net2026-07-172026-10-07hn
Kimi K3 Intelligence, Performance and Price Analysisartificialanalysis.ai2026-07-162026-10-07hn
Kimi K3 is now #1 in the Front end Code Arena with 1679 pts, surpassing Fable 5twitter.com2026-07-162026-10-07hn
Kimi K3: Open Frontier Intelligencekimi.com official2026-07-162026-10-07hn, exa

Similar models

3 models

Similar models

  • Inkling · 952B
    same size band (XXL), also MoE, same task (image-text-to-text), fewer providers, permissive license
  • Qwen3.8-2.4T-A95B · 2.45T
    same size band (XXL), also MoE, fewer providers
  • DeepSeek-V4-Pro · 1.60T
    same size band (XXL), also MoE, fewer providers, permissive license

Sources

5 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:31 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 14:59 UTC · Last verified 2026-10-25

Follow Kimi-K3

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.