Models / moonshotai

Kimi K3

moonshotai · Open Weight

Recent snapshotData snapshot cf-1.3
Cheese Index68.1cf-1.3
Input / output$3 / $15USD / 1M tokens
Measured dimensions8 / 13Evidence
Context window1,048,576tokens · conditions unverified

Decision brief

Start by investigating gpqa and Multimodal workloads. These are observed strengths in the dataset; validate them on your own tasks.

Coverage counts measured dimensions, not confidence. Benchmark conditions and public source links may be missing. Missing: Mathematics, Instructions, Multilingual, swe, terminal.

View evidence ↓

Measured task dimensions

Coding
58.7
Agent workflows
67.6
Mathematics
—
Knowledge
67.2
Instructions
—
Multimodal
90.0
Multilingual
—
Reasoning
69.7
swe
—
terminal
—
gpqa
93.5
hle
56.0
mcp
84.2

Normalized dimensions. A dash means unknown, not zero.

Input / output

$3 / $15 · USD / 1M tokens

Listed dataset prices; provider terms and dates need verification. Zero without explicit free-price evidence is treated as unknown.

Cache, batch, tool fees and self-hosting costs: not verified.

Index variability stability

100.0 · 7 days of history

band

This is index variability, not API uptime or reliability.

Before deployment

  • Confirm provider endpoint, license and version.
  • Test your prompts, tool calls and failure recovery.
  • Verify prices and latency with real usage.

Evidence

These are source categories published by the dataset. Original public links, sample sizes and harness details remain incomplete.

Source categoryProvenance typeAs ofOriginal public link
Codingcore / official_json2026-11-05Not published
Agent workflowscore / official_json2026-11-05Not published
Long workflowshorizon / official_json2026-11-05Not published
arenacross / official_json2026-11-05Not published
depthcore+cross / official_json2026-11-05Not published
Published index inputs · cf-1.3
Coding49.7
Agent workflows67.6
Long workflows68.5
Overall—

The method includes overall, but this snapshot does not expose that input. The total cannot be fully reproduced from these displayed axes.

Method: median of available axes, with an 8% Ops blend when present. Source scores are displayed without recalculation or imputation.

Method →
Index history
As ofCheese Index
2026-10-3068.1
2026-10-3168.1
2026-11-0168.1
2026-11-0268.1
2026-11-0368.1
2026-11-0468.1
2026-11-0568.1

Only recorded dates are shown. Missing days are not filled.