MiMo-V2.5-Pro

Leaderboards

Where MiMo-V2.5-Pro 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 MiMo-V2.5-Pro has a published result on, out of 100%, with its rank among the models that took it.

• GPQA

  #160 / 256

  Graduate-level science questions written to be hard to look up.

  66.7%

  Self-reported

• MMLU-Pro

  #99 / 142

  Broad knowledge and reasoning across 14 subjects, ten answer choices each.

  68.5%

  Self-reported

• SWE-Bench Verified

  #19 / 120

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

  78.9%

  Self-reported

• Humanity's Last Exam

  #51 / 108

  Humanity's Last Exam (HLE) is a multi-modal academic benchmark with 2,500 questions across mathematics, humanities, and natural sciences, designed to test LLM capabilities at the frontier of human knowledge with unambiguous, verifiable solutions

  34.0%

  Self-reported · In the text score· In the vision score

• MMLU

  #16 / 103

  Massive Multitask Language Understanding benchmark testing knowledge across 57 diverse subjects including STEM, humanities, social sciences, and professional domains

  89.4%

  Self-reported

• MATH

  #11 / 71

  MATH dataset contains 12,500 challenging competition mathematics problems from AMC 10, AMC 12, AIME, and other mathematics competitions. Each problem includes full step-by-step solutions and spans multiple difficulty levels (1-5) across seven mathematical subjects including Prealgebra, Algebra, Number Theory, Counting and Probability, Geometry, Intermediate Algebra, and Precalculus.

  86.2%

  Self-reported

• LiveCodeBench v6

  #64 / 67

  LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code. It continuously collects new problems from programming contests (LeetCode, AtCoder, CodeForces) and evaluates four different scenarios: code generation, self-repair, code execution, and test output prediction. Problems are annotated with release dates to enable evaluation on unseen problems released after a model's training cutoff.

  39.6%

  Self-reported

• SWE-Bench Pro

  #31 / 61

  Harder, longer software engineering tasks from real codebases.

  57.2%

  Self-reported · In the coding score

• Terminal-Bench 2.0

  #13 / 54

  Getting real work done in a command-line terminal.

  68.4%

  Self-reported

• MMLU-Redux

  #15 / 50

  An improved version of the MMLU benchmark featuring manually re-annotated questions to identify and correct errors in the original dataset. Provides more reliable evaluation metrics for language models by addressing dataset quality issues found in the original MMLU.

  92.8%

  Self-reported

• GSM8k

  #1 / 48

  Grade School Math 8K, a dataset of 8.5K high-quality linguistically diverse grade school math word problems requiring multi-step reasoning and elementary arithmetic operations.

  99.6%

  Self-reported

• ARC-C

  #1 / 35

  The AI2 Reasoning Challenge (ARC) Challenge Set is a multiple-choice question-answering benchmark containing grade-school level science questions that require advanced reasoning capabilities. ARC-C specifically contains questions that were answered incorrectly by both retrieval-based and word co-occurrence algorithms, making it a particularly challenging subset designed to test commonsense reasoning abilities in AI systems.

  97.2%

  Self-reported

• DROP

  #4 / 30

  DROP (Discrete Reasoning Over Paragraphs) is a reading comprehension benchmark requiring discrete reasoning over paragraph content. It contains crowdsourced, adversarially-created questions that require resolving references and performing discrete operations like addition, counting, or sorting, demanding comprehensive paragraph understanding beyond paraphrase-and-entity-typing shortcuts.

  86.3%

  Self-reported

• Finance Agent v2

  #17 / 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.

  41.5%

  Self-reported

• HellaSwag

  #5 / 28

  A challenging commonsense natural language inference dataset that uses Adversarial Filtering to create questions trivial for humans (>95% accuracy) but difficult for state-of-the-art models, requiring completion of sentence endings based on physical situations and everyday activities

  89.8%

  Self-reported

• Winogrande

  #3 / 23

  WinoGrande: An Adversarial Winograd Schema Challenge at Scale. A large-scale dataset of 44,000 pronoun resolution problems designed to test machine commonsense reasoning. Uses adversarial filtering to reduce spurious biases and provides a more robust evaluation of whether AI systems truly understand commonsense or exploit statistical shortcuts. Current best AI methods achieve 59.4-79.1% accuracy, significantly below human performance of 94.0%.

  85.6%

  Self-reported

• C-Eval

  #6 / 20

  C-Eval is a comprehensive Chinese evaluation suite designed to assess advanced knowledge and reasoning abilities of foundation models in a Chinese context. It comprises 13,948 multiple-choice questions across 52 diverse disciplines spanning humanities, science, and engineering, with four difficulty levels: middle school, high school, college, and professional. The benchmark includes C-Eval Hard, a subset of very challenging subjects requiring advanced reasoning abilities.

  91.5%

  Self-reported

• TriviaQA

  #3 / 18

  A large-scale reading comprehension dataset containing over 650K question-answer-evidence triples. TriviaQA includes 95K question-answer pairs authored by trivia enthusiasts and independently gathered evidence documents (six per question on average) that provide high quality distant supervision for answering the questions. The dataset features relatively complex, compositional questions with considerable syntactic and lexical variability, requiring cross-sentence reasoning to find answers.

  81.3%

  Self-reported

• Claw-Eval

  #6 / 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.

  64.0%

  Self-reported

• BBH

  #2 / 13

  Big-Bench Hard (BBH) is a suite of 23 challenging tasks selected from BIG-Bench for which prior language model evaluations did not outperform the average human-rater. These tasks require multi-step reasoning across diverse domains including arithmetic, logical reasoning, reading comprehension, and commonsense reasoning. The benchmark was designed to test capabilities believed to be beyond current language models and focuses on evaluating complex reasoning skills including temporal understanding, spatial reasoning, causal understanding, and deductive logical reasoning.

  88.4%

  Self-reported

• HumanEval+

  #7 / 10

  Enhanced version of HumanEval that extends the original test cases by 80x using EvalPlus framework for rigorous evaluation of LLM-synthesized code functional correctness, detecting previously undetected wrong code

  75.6%

  Self-reported

• Global-MMLU

  #4 / 8

  A comprehensive multilingual benchmark covering 42 languages that addresses cultural and linguistic biases in evaluation, with improved translation quality and culturally sensitive question subsets.

  83.6%

  Self-reported

• TAU3-Bench

  #1 / 8

  TAU3-Bench is a benchmark for evaluating general-purpose agent capabilities, testing models on multi-turn interactions with simulated user models, retrieval, and complex decision-making scenarios.

  72.9%

  Self-reported

• CMMLU

  #1 / 6

  CMMLU (Chinese Massive Multitask Language Understanding) is a comprehensive Chinese benchmark that evaluates the knowledge and reasoning capabilities of large language models across 67 different subject topics. The benchmark covers natural sciences, social sciences, engineering, and humanities with multiple-choice questions ranging from basic to advanced professional levels.

  90.2%

  Self-reported

• WildClawBench

  #6 / 6

  WildClawBench is an agentic coding benchmark from InternLM/Claw-Eval that reports overall model performance on real-world tool-using development tasks.

  43.0%

  Self-reported

• MBPP+

  #1 / 4

  MBPP+ is an enhanced version of MBPP (Mostly Basic Python Problems) with significantly more test cases (35x) for more rigorous evaluation. MBPP is a benchmark of 974 crowd-sourced Python programming problems designed to be solvable by entry-level programmers, covering programming fundamentals and standard library functionality.

  74.1%

  Self-reported

• MiMo Coding Bench

  #1 / 4

  MiMo Coding Bench evaluates coding-agent capabilities on software engineering tasks reported with the MiMo model family.

  73.7%

  Self-reported

• GraphWalks

  #3 / 3

  GraphWalks is a synthetic multi-hop long-context reasoning benchmark in which a model is given an edge-list representation of a graph and must traverse it to find neighboring nodes (via breadth-first search) or parent nodes for a given start node. Performance is reported as F1 of the model-predicted answer set versus the ground truth.

  62.0%

  Self-reported

• AIME

  #2 / 2

  American Invitational Mathematics Examination (AIME) benchmark for evaluating mathematical reasoning capabilities of large language models. Contains 30 challenging mathematical problems from AIME 2024 competition that require multi-step reasoning and advanced mathematical insight. Each problem has an integer answer between 000-999.

  37.3%

  Unverified

• SWE-bench Verified (Agentless)

  #2 / 2

  A human-validated subset of SWE-bench that evaluates language models' ability to resolve real-world GitHub issues using an agentless approach. The benchmark tests models on software engineering problems requiring understanding and coordinating changes across multiple functions, classes, and files simultaneously.

  35.7%

  Unverified

• FrontierSWE (Impl.)

  #1 / 1

  FrontierSWE (Impl.) evaluates software engineering implementation ability and reports model ranking on implementation tasks. Lower rank is better.

  340.0%

  Unverified