Kuaishou

KAT-Coder V2.5

A coding-focused agentic model trained for autonomous work in executable repositories.

codingagents

Model record

KAT-Coder V2.5 model overview

KAT-Coder V2.5 is a api model from Kuaishou. It was released on July 6, 2026. Its strongest case is repository-level software engineering, while buyers should account for no developer-owned model card was available.

KAT-Coder V2.5 benchmark snapshot

EvaluationReported scoreWhat it probes
MMLU-ProNot reportedBroad knowledge and multi-step reasoning
GPQA DiamondNot reportedGraduate-level science reasoning
SWE-bench VerifiedNot reportedVerified real-repository issue resolution
LiveCodeBenchNot reportedContamination-aware competitive programming
SWE-Bench ProNot reportedLonger, harder professional software tasks
Artificial Analysis Intelligence IndexNot reportedComposite third-party capability index; version matters

Scores are percentages reported by model creators or benchmark maintainers under varying settings. A blank is preferable to an inferred result. See our methodology.

KAT-Coder V2.5 strengths

  • Repository-level software engineering
  • Tool use
  • Executable-environment post-training

KAT-Coder V2.5 limitations

  • No developer-owned model card was available
  • Deployment limits, license, prices, and scale unavailable
Editorial take: Benchmark rank should narrow a shortlist, not close a purchase. Run a private evaluation with representative prompts, failure cases, latency targets, and total token costs.

Source record

Specifications and scores are linked to the best source located during review. Provider-reported results are not presented as third-party lab reproductions.

Source limitation: A developer-owned model card was unavailable. This author technical report is used under the explicit audit exception; provider-reported benchmarks are not imported into the shared taxonomy.

Read KAT-Coder-V2.5 author technical report