MiniMax M3

Leaderboards

Where MiniMax M3 sits among the top 20 in each category.

By task

Scored from votes on each kind of prompt alone, on the same 1,000-point scale.

Test results

Every test MiniMax M3 has a published result on, out of 100%, with its rank among the models that took it.

• SWE-Bench Verified

  #11 / 120

  Fixing real bugs in real GitHub projects, checked by the projects' own tests.

  80.5%

  Self-reported

• MMMU-Pro

  #22 / 73

  College-level questions that need both an image and text to answer.

  78.1%

  Self-reported · In the vision score

• BrowseComp

  #17 / 68

  Finding hard-to-locate facts by browsing the web.

  83.5%

  Self-reported · In the agents score

• SWE-Bench Pro

  #22 / 61

  Harder, longer software engineering tasks from real codebases.

  59.0%

  Self-reported · In the coding score

• Terminal-Bench 2.1

  #32 / 46

  Terminal-Bench 2.1 is an updated release of the Terminal-Bench benchmark that tests AI agents' ability to operate a computer via the terminal. It evaluates how well models handle real-world, end-to-end tasks autonomously, including compiling code, training models, setting up servers, system administration, data science workflows, and security tasks.

  66.0%

  Self-reported

• LiveBench

  #31 / 38

  LiveBench is a challenging, contamination-limited LLM benchmark that addresses test set contamination by releasing new questions monthly based on recently-released datasets, arXiv papers, news articles, and IMDb movie synopses. It comprises tasks across math, coding, reasoning, language, instruction following, and data analysis with verifiable, objective ground-truth answers.

  70.0%

  Self-reported · In the text score

• MCP Atlas

  #18 / 38

  MCP Atlas is a benchmark for evaluating AI models on scaled tool use capabilities, measuring how well models can coordinate and utilize multiple tools across complex multi-step tasks.

  74.2%

  Self-reported

• Finance Agent v2

  #11 / 28

  Finance Agent v2 is an agentic financial-analysis benchmark from Vals that evaluates models on real-world finance workflows, measuring their ability to retrieve and reason over financial documents, perform multi-step calculations, and produce accurate analyses.

  48.3%

  Self-reported

• OSWorld-Verified

  #20 / 28

  OSWorld-Verified is a verified subset of OSWorld, a scalable real computer environment for multimodal agents supporting task setup, execution-based evaluation, and interactive learning across Ubuntu, Windows, and macOS.

  70.1%

  Self-reported

• VideoMMMU

  #7 / 28

  Video-MMMU evaluates Large Multimodal Models' ability to acquire knowledge from expert-level professional videos across six disciplines through three cognitive stages: perception, comprehension, and adaptation. Contains 300 videos and 900 human-annotated questions spanning Art, Business, Science, Medicine, Humanities, and Engineering.

  84.6%

  Self-reported

• NL2Repo

  #20 / 25

  NL2Repo evaluates long-horizon coding capabilities including repository-level understanding, where models must generate or modify code across entire repositories from natural language specifications.

  42.1%

  Self-reported · In the coding score

• FrontierCode 1.1

  #22 / 23

  FrontierCode 1.1 evaluates whether coding-agent changes are mergeable, using unit tests, maintainer-defined rubrics, and verifiers. Runs flagged for unfair internet use receive a zero score.

  14.7%

  Self-reported · In the coding score

• OmniDocBench 1.5

  #1 / 20

  OmniDocBench 1.5 is a comprehensive benchmark for evaluating multimodal large language models on document understanding tasks, including OCR, document parsing, information extraction, and visual question answering across diverse document types. Lower Overall Edit Distance scores are better.

  91.6%

  Self-reported

• Video-MME

  #6 / 17

  Video-MME is the first-ever comprehensive evaluation benchmark of Multi-modal Large Language Models (MLLMs) in video analysis. It features 900 videos totaling 254 hours with 2,700 human-annotated question-answer pairs across 6 primary visual domains (Knowledge, Film & Television, Sports Competition, Life Record, Multilingual, and others) and 30 subfields. The benchmark evaluates models across diverse temporal dimensions (11 seconds to 1 hour), integrates multi-modal inputs including video frames, subtitles, and audio, and uses rigorous manual labeling by expert annotators for precise assessment.

  85.4%

  Self-reported

• Claw-Eval

  #3 / 14

  Claw-Eval tests real-world agentic task completion across complex multi-step scenarios, evaluating a model's ability to use tools, navigate environments, and complete end-to-end tasks autonomously.

  74.5%

  Self-reported

• OfficeQA Pro

  #12 / 12

  OfficeQA Pro evaluates AI models on professional knowledge-work questions and tasks drawn from real office workflows, including document analysis, spreadsheet reasoning, and information synthesis across business domains.

  45.1%

  Self-reported

• APEX-Agents

  #10 / 11

  APEX-Agents is a benchmark evaluating AI agents on long horizon professional tasks that require sustained reasoning, planning, and execution across complex multi-step workflows.

  27.7%

  Self-reported

• PostTrainBench

  #4 / 9

  PostTrainBench evaluates a model's ability to autonomously post-train base models. Given pretrain-only base models, the agent must complete the full pipeline of data synthesis, training, evaluation, and iteration within a time budget, scored across downstream benchmarks such as AIME2025, BFCL, GPQA Main, GSM8K, and HumanEval.

  37.1%

  Self-reported

• SWE Atlas - Codebase QnA

  #5 / 5

  SWE Atlas - Codebase QnA evaluates a model's ability to answer questions about real codebases, measuring repository-level comprehension and the ability to reason about code structure, behavior, and intent across an entire project.

  37.9%

  Self-reported

• SpreadSheetBench-v1

  #1 / 4

  SpreadSheetBench-v1 evaluates office automation agents on spreadsheet reasoning and manipulation tasks, measuring the ability to analyze, transform, and operate on spreadsheet data through tools.

  89.3%

  Self-reported

• BankerToolBench

  #3 / 3

  BankerToolBench is a public benchmark that evaluates models on banking and finance tool-use tasks. Models are scored against dataset rubrics, measuring their ability to correctly invoke tools and complete multi-step financial workflows.

  76.1%

  Self-reported

• CL-bench

  #2 / 3

  CL-bench is an open-source benchmark with its own data and rubrics for evaluating models on coding and agentic tasks, scored using a setup fully aligned with the official procedure.

  20.5%

  Self-reported

• DRACO

  #2 / 3

  DRACO is a deep research benchmark that evaluates an agent's ability to gather, synthesize, and reason over information to answer complex research questions. Scores are based on official rubrics per question, with the final score being the average across all questions.

  73.2%

  Self-reported

• PaperBench

  #3 / 3

  PaperBench is a benchmark for evaluating AI agents on their ability to replicate research papers. It tests models on complex, multi-step workflows involving code implementation, experimentation, and reproducing scientific results from academic publications.

  52.6%

  Self-reported

• USAMO 2026

  #1 / 3

  USAMO 2026 evaluates models on the six problems from the 2026 United States of America Mathematical Olympiad, requiring rigorous proof-based reasoning. Following the MathArena methodology, proofs are graded against human expert rubrics by dual strong judge models, with the minimum of the two scores taken as final. Maximum score is 42 points.

  85.7%

  Self-reported

• IMO 2025

  #1 / 2

  IMO 2025 evaluates models on the six problems from the 2025 International Mathematical Olympiad, requiring rigorous proof-based reasoning. Following the MathArena methodology, proofs are graded against human expert rubrics by dual strong judge models, with the minimum of the two scores taken as final. Maximum score is 42 points.

  83.3%

  Unverified

• LOCA-Bench (256k)

  #2 / 2

  LOCA-Bench is a long-context agentic benchmark. The 256k variant evaluates agents using the official ReAct mode with an environment description length of 256k tokens, measuring how well models reason and act over very long contexts.

  49.3%

  Unverified

• GDPval-Rubrics

  #1 / 1

  GDPval-Rubrics evaluates AI model performance on economically valuable knowledge work tasks drawn from the public GDPval dataset. It uses pointwise scoring based on public rubrics, with the environment aligned to the GDPval-AA scaffolding.

  74.8%

  Unverified

• KernelBench Hard

  #1 / 1

  KernelBench Hard evaluates agentic GPU kernel optimization on the hardest problem set. Each question is scored by the agent's submitted operator TFLOPs relative to the theoretical peak of the current hardware, with the benchmark score being the average across all questions.

  28.8%

  Unverified

• LiveSQLBench

  #1 / 1

  LiveSQLBench evaluates models on generating correct SQL queries against live PostgreSQL databases. The LiveSQLBench-Base-Full v1 dataset contains 600 questions across 22 PostgreSQL databases, testing real-world database reasoning and query generation.

  40.2%

  Unverified

• SVG-Bench

  #1 / 1

  SVG-Bench is an internal benchmark that comprehensively evaluates SVG generation performance. It accepts text and image inputs across build-from-scratch and edit-based tasks, using a VLM to verify rendering accuracy of the generated outputs.

  63.7%

  Unverified

• SWE Atlas - Test Writing

  #1 / 1

  SWE Atlas - Test Writing evaluates a model's ability to author meaningful tests for real-world software projects, measuring how well agents can understand code and produce correct, useful test coverage.

  30.8%

  Unverified

• SWE-fficiency

  #1 / 1

  SWE-fficiency is an open-source benchmark and workflow that evaluates language models on optimizing the runtime efficiency of real-world software engineering tasks, measuring how well agents can improve code performance autonomously.

  34.8%

  Unverified

• VIBE-V2

  #1 / 1

  VIBE-V2 is an internal benchmark covering pure front-end and full-stack Web, Android, and iOS projects with build-from-scratch tasks. It uses an Agent-as-a-Verifier paradigm to automatically verify program interaction logic and visual output, scoring models through a unified pipeline that includes a requirement set, containerized deployment, and a dynamic interaction environment.

  50.1%

  Unverified

• YC-Bench

  #1 / 1

  YC-Bench evaluates agents on long-horizon, open-ended business and investment decision-making. The reported metric is the final assets (fund value, in US dollars) accumulated by the agent over the course of the simulation.

  21.0%

  Unverified