m/.eval_results/minimax-m3.yaml

# Evaluation results for MiniMaxAI/MiniMax-M3
# Extracted from the model card benchmark graph (figures/benchmark.jpeg)
# https://huggingface.co/MiniMaxAI/MiniMax-M3
# Paper: https://arxiv.org/abs/2606.13392

# ---------------------------------------------------------------------------
# Coding
# ---------------------------------------------------------------------------

# SWE-Bench Verified - 80.5
- dataset:
    id: SWE-bench/SWE-bench_Verified
    task_id: swe_bench_%_resolved
  value: 80.5
  source:
    url: https://huggingface.co/MiniMaxAI/MiniMax-M3
    name: MiniMax-M3 model card
  notes: "Evaluated on internal infrastructure using Claude Code as the scaffolding. Each test was run 4 times and the average was taken."

# SWE-Bench Pro - 59.0
- dataset:
    id: ScaleAI/SWE-bench_Pro
    task_id: SWE_Bench_Pro
  value: 59.0
  source:
    url: https://huggingface.co/MiniMaxAI/MiniMax-M3
    name: MiniMax-M3 model card
  notes: "Evaluated on internal infrastructure using Claude Code as the scaffolding. Testing logic is aligned with the official evaluation."


# ---------------------------------------------------------------------------
# Multimodal
# ---------------------------------------------------------------------------

# MMMU-Pro - 78.1
# MMMU-Pro defines three tasks: mmmu_pro_vision, mmmu_pro_standard_4_options,
# mmmu_pro_standard_10_options. The model card reports a single "MMMU-Pro"
# score without specifying the exact variant. We map it to the standard
# 10-options task as the most common updated benchmark configuration.
- dataset:
    id: MMMU/MMMU_Pro
    task_id: mmmu_pro_standard_10_options
  value: 78.1
  source:
    url: https://huggingface.co/MiniMaxAI/MiniMax-M3
    name: MiniMax-M3 model card
  notes: "MMMU-Pro score extracted from the model card benchmark graph. The exact task variant (vision, standard 4-options, or standard 10-options) is not explicitly stated."

# Video-MME (w/ sub) - 85.4
# Mapped to Video-MME-v2, the registered successor benchmark on the Hub.
- dataset:
    id: MME-Benchmarks/Video-MME-v2
    task_id: video-mme-v2
  value: 85.4
  source:
    url: https://huggingface.co/MiniMaxAI/MiniMax-M3
    name: MiniMax-M3 model card
  notes: "Model card reports 'VideoMME (w/ sub)'. Mapped to the closest registered benchmark on the Hub, Video-MME-v2."

# ---------------------------------------------------------------------------
# Cowork
# ---------------------------------------------------------------------------

# Claw-Eval - 74.5
# Claw-Eval defines three tasks: general, multimodal, multi_turn. The model card
# reports a single overall score, so it is mapped to the 'general' task.
- dataset:
    id: claw-eval/Claw-Eval
    task_id: general
  value: 74.5
  source:
    url: https://huggingface.co/MiniMaxAI/MiniMax-M3
    name: MiniMax-M3 model card
  notes: "Model card reports a single 'Claw-Eval' score. Mapped to the 'general' task (overall); the exact task split is not specified."

# Apex-Agents - 27.7
- dataset:
    id: mercor/apex-agents
    task_id: apex-agents
  value: 27.7
  source:
    url: https://huggingface.co/MiniMaxAI/MiniMax-M3
    name: MiniMax-M3 model card
  notes: "Evaluated on the apex-agents benchmark."

# YC-Bench - 2.1M (final assets in fund, monetary metric)
- dataset:
    id: collinear-ai/yc-bench
    task_id: medium
  value: 2100000
  source:
    url: https://huggingface.co/MiniMaxAI/MiniMax-M3
    name: MiniMax-M3 model card
  notes: "Model card reports 2.1M (monetary value, final assets fund). The benchmark's 'medium' task is used as the overall evaluation. Metric is monetary, not percentage-based."