stealthmodels.

MODEL BENCHMARKS

Kimi K3

Open-weight multimodal reasoning at a 2.8T-parameter scale, with a million-token context and 104B parameters active per token.

BENCHMARK BY BENCHMARK

Kimi K3 benchmark scores

StealthMark
50.6/100Overall rank 19/58
General Intelligence
44.4/100
Visual Intelligence
61.4/100

Kimi K3Other modelsMedian

Reasoning efficiency

How many reasoning tokens a model needs for the quality it delivers.

Single-step reasoningHow much it reasons to understand a taskRank 27/56
Multi-step reasoningHow much it keeps reasoning as a task goes onRank 41/54

Cost efficiency

How much Performance you get for your money.

Rank 45/52

DATASHEET

Kimi K3 specs and pricing

A good fit forComplex research and agent workflows combining long documents and images. Hosted access avoids the large memory footprint of self-hosting.

Kimi K3Other modelsMedian

PriceInput / output per 1M tokens
$0.239 / $10.80Cheaper than 7 of 46 others
Context windowTokens in one request
1,048,576Joint largest with 21 others
Output speedTokens per second, our measurement
115 tokens/sFaster than 20 of 51 others
Developer
Moonshot AI
Released
16 Jul 2026
Family
Kimi K3
Open model
Yes
Local model
No
Origin
China
Parameters
2.8T
Active parameters
104B
Max output
1,048,576 tokens
Input → output
Text, image → text
Architecture
Mixture-of-Experts with KDA + Gated MLA attention, 896 experts and native vision encoder

Self-host Kimi K3

Hardware
Rack-scaleEstimated
Download size
649 GB (~1-bit) to 1,510 GB (~4-bit)

Official Hugging Face weightsUnsloth GGUF

All prices and providers
  • OpenRouterObserved 2026-10-23
    Input $0.239Output $10.8Cached read $0.21

    per 1M tokens

    moonshotai/kimi-k3
Deployment details
  • Unsloth UD-IQ1_M GGUF649 GB · ~1-bit
  • Unsloth UD-Q4_K_XL GGUF1,510 GB · ~4-bit
Measurements and sources9

Related models

  1. Kimi K3.1 Pro56.7Compare Kimi K3 and Kimi K3.1 Pro
  2. Kimi K350.6This model

GAME BENCHMARKS

Watch Kimi K3 play

Foresight

Anticipating threats and planning ahead

52.7/100Rank 19/54
Survived 0 of 3
Foresight: complete playing field
Level 311 ptsTurn 17 / 20

Points per level

  1. Level 1Second Thoughts17 ptsBest 27Out of lives
  2. Level 2Pressure Pot10 ptsBest 13Out of lives
  3. Level 3Hide or Flight11 ptsBest 19Out of lives

Collect coins while avoiding ghosts. Ghosts move before the player each turn, and collisions cost a life. 3 levels in sequence, up to 66 turns in total. Earlier moves and outcomes carry forward, testing learning across levels.

Latch

Cooperative planning and learning from experience

57.7/100Rank 27/55
Cleared 3 of 5
Latch: complete playing field
Level 514.2 ptsTurn 4 / 40

Points per level

  1. Level 1Hold the door5.2 ptsBest 5.2Cleared
  2. Level 2Crossfeed9.6 ptsBest 10.2Cleared
  3. Level 3Relay vault18.2 ptsBest 19.2ClearedFinished early · 3 beacons left
  4. Level 4Passing bay14.2 ptsBest 18.2Finished early
  5. Level 5Second thoughts14.2 ptsBest 20.8Finished early

Coordinate two robots to push numbered cores onto matching sockets. Pressure plates hold doors open; moving either robot uses one turn. 5 levels in sequence, up to 152 turns in total. Earlier moves and outcomes carry forward, testing learning across levels.

Kimi K3 benchmark results ↗

StealthMark score

StealthMark measures how well a model’s abilities carry over to original, demanding tasks it could not have trained for: physical and spatial reasoning, planning across many steps, and precise visual work. How StealthMark works explains every benchmark, what it tests and why it is hard.

General Intelligence

The ability to reason, plan, solve problems and understand how things relate and behave. A high score means it thinks hard problems through, from careful analysis to plans that span many steps.

Visual Intelligence

The ability to understand visual intent and turn ideas, descriptions and plans into coherent, effective visuals. A high score means design ideas come out the way you meant them, in illustrations, icons, diagrams or interface visuals.

Overall score

Overall shows how well a model’s abilities carry over to original, demanding tasks, weighting every benchmark into one score. Estimated means the total includes an estimate for one missing scene.

Judging

Scores reflect how well each result meets the benchmark requirements. Game scores come from the recorded play: the points earned on each level and how each level ended. How StealthMark keeps scores honest.

Games

Each game is a sequence of levels, and earlier moves carry into later levels. A level that was not played is shown as not played and never counts as zero. Play every level of Latch or Foresight yourself to get a score calculated like the models’.

Reasoning efficiency

Good results with little thinking score high. On each task, a model’s result and reasoning tokens are compared with every other model and combined with the Balanced Integration Score (Liesefeld & Janczyk, 2019); 50 is average. Single step shows how much reasoning a model needs to understand and solve one problem, multi-step how efficiently it keeps working on a task by building on its earlier reasoning, as in agentic workflows. Not part of StealthMark.

Cost efficiency

How much StealthMark a model delivers for its money, in StealthMark points per dollar: its overall score divided by what one benchmark task costs it on average. The cost is what our runs were charged, or their tokens at the model’s list price; benchmarks a model has not run yet are estimated from its other costs. Like StealthMark, it is absolute and compares with no other model. Not part of StealthMark.

BENCHMARK RESULTS

Compare results

This link opens the same results in the same order.

StealthMark benchmark results

Reset benchmark preferences?

Return to the default view, clear comparison selections and restore the advertisement.