MidMarketBench
Rank 1completeObserved OpenRouter runProvisional tie

Kimi K3

Route caveat

Read this result with its endpoint.

Kimi K3 via Moonshot AI, INT4, OpenRouter. Mandatory max reasoning; full weights were not yet released on 18 July 2026.

Model and run

Provider
Moonshot AI
Version
kimi-k3
Release
Not published
Context
1,048,576 tokens
Weights
Not published
Run status
complete
Candidate samples
2/2
Scored samples
2/2
Overall
96.3
Observed range
95.6 to 96.9
Provider-reported attributed cost
$0.894
Candidate attempts
$0.613
Judge attempts
$0.282
Median candidate latency
6.4 minutes

Model source: OpenRouter catalogue

Attributed costs include every OpenRouter-reported candidate and judge attempt for this model, including retries. The run total is authoritative for settled key spend.

OpenRouter provenance

Requested model and returned route.

Requested model
moonshotai/kimi-k3
Routed provider
Moonshot AI
Endpoint tag
moonshotai/int4
Returned model
moonshotai/kimi-k3
Quantisation
INT4
Pinned route price
moonshotai/int4: $3.000 input / $15.000 output per 1M tokens

Run 2026-07-18-final / 2026-07-18 / Closed-book

Dimension shape

Where the observed score comes from.

Overall combines deterministic task checks with blinded, calibrated, cross-family judgements of the IC note.

Grounding85.4
Commercial judgement98.7
Scepticism97.7
Numerical sanity100.0
Risk discovery96.6
Question generation96.1
European context100.0
Output usefulness99.7

Complete field

Position among ranked two-sample runs.

6498

Interpretation limit

One case, 2 scored samples.

Directional mini benchmark: one fresh synthetic case and two samples per model. It is not a universal ranking of model intelligence.