MiniMax-M2: a Mini model built for Max coding & agentic workflows

TL;DR

MiniMax-M2 is an open-weights (MIT-license) sparse-MoE LLM with 230B total / 10B active parameters, positioned by MiniMax as a coding-and-agentic-workflows specialist that hits the small-active-parameter sweet spot the rest of the open-MoE cluster is converging on. The README’s headline claims are that MiniMax-M2 is #1 among open-source models on the Artificial Analysis composite score and competitive with Claude Sonnet 4.5 / GPT-5 (thinking) on end-to-end agentic-coding benchmarks (SWE-bench Verified 69.4, Multi-SWE-Bench 36.2, Terminal-Bench 46.3, ArtifactsBench 66.8) and on long-horizon search/browse tasks (BrowseComp 44, BrowseComp-zh 48.5, GAIA-text 75.7, τ²-Bench 77.2). It is an interleaved thinking model using <think>…</think> segments that must be preserved across multi-turn history. Day-0 serving on SGLang, vLLM, and MLX-LM; weights at huggingface.co/MiniMaxAI/MiniMax-M2.

Key claims

Architecture is sparse MoE at 230B total / 10B active parameters [§Meet MiniMax-M2].
Composite “AA Intelligence” score 61, reported as #1 among open-source models on the Artificial Analysis methodology, vs Claude Sonnet 4 57, Claude Sonnet 4.5 63, Gemini 2.5 Pro 60, GPT-5 (thinking) 69, GLM-4.6 56, Kimi K2 0905 50, DeepSeek-V3.2 57 [Intelligence Benchmarks table].
SWE-bench Verified 69.4 — between Claude Sonnet 4 72.7 and DeepSeek-V3.2 67.8 — evaluated with an R2E-Gym-style scaffold on top of OpenHands, 128k context, 100 max steps, no test-time scaling, all git-related content stripped so the agent only sees code at issue time [Coding & Agentic Benchmarks table, Notes].
Multi-SWE-Bench 36.2 and SWE-bench Multilingual 56.5, averaged over 8 runs with the claude-code CLI (300 max steps) as the scaffold [Coding & Agentic Benchmarks table, Notes].
Terminal-Bench 46.3 (mean pass rate over 8 runs with official claude-code from terminal-bench commit 94bf692), beating Claude Sonnet 4 36.4 and Gemini 2.5 Pro 25.3 and matching Claude Sonnet 4.5 50.0 closely [Coding & Agentic Benchmarks table, Notes].
BrowseComp 44.0 and BrowseComp-zh 48.5, evaluated with the WebExplorer agent framework — substantially above Claude Sonnet 4 12.2 / Claude Sonnet 4.5 19.6 / Gemini 2.5 Pro 9.9, and competitive with GLM-4.6 45.1 / DeepSeek-V3.2 40.1 [Coding & Agentic Benchmarks table, Notes].
GAIA text-only 75.7 on the 103-sample text-only validation subset, xbench-DeepSearch 72, τ²-Bench 77.2, HLE-with-tools 31.8 [Coding & Agentic Benchmarks table].
The 10B-activation design is motivated explicitly as a plan→act→verify-loop primitive — faster feedback cycles in compile-run-test and browse-retrieve-cite chains, more concurrent runs at fixed budget, simpler capacity planning [§Why activation size matters].
Interleaved thinking via <think>…</think> segments that must be preserved in historical messages across multi-turn turns; recommended inference parameters temperature=1.0, top_p=0.95, top_k=40 [§Local Deployment Guide, §Inference Parameters].
Released openly with MIT license, weights on HuggingFace (MiniMaxAI/MiniMax-M2), code on GitHub, and day-0 deployment guides for SGLang, vLLM, and MLX-LM [§Local Deployment Guide].
API and the MiniMax Agent product (built on M2) are publicly available and free for a limited time on platform.minimax.io and agent.minimax.io [§How to Use].

Method

The README is a release post, not a technical report — no training recipe, no data-mix details, no architecture diagram beyond “230B total / 10B active sparse MoE”. Architecturally, the 230B-total / 10B-active geometry sits between GLM-4.6 (32B-active) and DeepSeek-V3.2 (37B-active) but at a substantially lower activation ratio (~4.3%) — close to Qwen3-Coder-Next 80B-total / 3B-active (~3.8%). The README characterizes the model as built for “elite performance in coding and agentic tasks, all while maintaining powerful general intelligence” and explicitly markets the activation count as the load-bearing design choice for agent loops.

Inference deployment is documented across SGLang, vLLM, and MLX-LM with separate deployment guides linked from the HuggingFace model card. A separate Tool Calling Guide document is referenced. The release does not disclose pretraining tokens, post-training stages, MoE topology (expert count, top-k), or RL recipe; future tech-report material (if any) is not yet published.

Results

AA-Intelligence composite: M2 61, GPT-5 (thinking) 69, Claude Sonnet 4.5 63, Gemini 2.5 Pro 60, Claude Sonnet 4 57, DeepSeek-V3.2 57, GLM-4.6 56, Kimi K2 0905 50 [Intelligence Benchmarks table].
SWE-bench Verified: M2 69.4, Claude Sonnet 4.5 77.2, GPT-5 (thinking) 74.9, Claude Sonnet 4 72.7, Kimi K2 0905 69.2, GLM-4.6 68.0, DeepSeek-V3.2 67.8, Gemini 2.5 Pro 63.8 [Coding & Agentic Benchmarks table].
Terminal-Bench: M2 46.3, Claude Sonnet 4.5 50.0, Kimi K2 0905 44.5, GPT-5 (thinking) 43.8, GLM-4.6 40.5, DeepSeek-V3.2 37.7, Claude Sonnet 4 36.4, Gemini 2.5 Pro 25.3 [Coding & Agentic Benchmarks table].
ArtifactsBench (Gemini-2.5-Pro judge, three-run mean): M2 66.8, GPT-5 (thinking) 73.0, Claude Sonnet 4.5 61.5, GLM-4.6 59.8, Claude Sonnet 4 57.3, Gemini 2.5 Pro 57.7, DeepSeek-V3.2 55.8, Kimi K2 0905 54.2 [Coding & Agentic Benchmarks table].
BrowseComp: M2 44.0, GPT-5 (thinking) 54.9, GLM-4.6 45.1, DeepSeek-V3.2 40.1, Claude Sonnet 4.5 19.6, Kimi K2 0905 14.1, Claude Sonnet 4 12.2, Gemini 2.5 Pro 9.9 [Coding & Agentic Benchmarks table].
GAIA text-only: M2 75.7, GPT-5 (thinking) 76.4, GLM-4.6 71.9, Claude Sonnet 4.5 71.2, Claude Sonnet 4 68.3, DeepSeek-V3.2 63.5, Gemini 2.5 Pro / Kimi K2 0905 60.2 [Coding & Agentic Benchmarks table].
τ²-Bench (extended thinking with tool use, GPT-4.1 user simulator): M2 77.2, Claude Sonnet 4.5 84.7, GPT-5 (thinking) 80.1, GLM-4.6 75.9, Kimi K2 0905 70.3, DeepSeek-V3.2 66.7, Claude Sonnet 4 65.5, Gemini 2.5 Pro 59.2 [Coding & Agentic Benchmarks table].
FinSearchComp-global: M2 65.5, GPT-5 (thinking) 63.9, Claude Sonnet 4.5 60.8, Gemini 2.5 Pro 42.6, Claude Sonnet 4 42.0, Kimi K2 0905 29.5, GLM-4.6 29.2, DeepSeek-V3.2 26.2 [Coding & Agentic Benchmarks table].

Why it’s interesting

MiniMax-M2 is the named MiniMax M2.x family member that Agentic Software Engineering was already tracking abstractly from third-party Pareto plots — Qwen3-Coder-Next’s own technical report lists “MiniMax M2.1 10B-active” on the (active parameters, SWE-Bench Pro) frontier (Qwen3-Coder-Next Technical Report Fig. 1), and Step 3.5 Flash’s peer benchmark table already normalizes inference cost against “MiniMax M2.1” at 3.9× (Step 3.5 Flash Benchmark table). M2 also extends the lineage from MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention (456B-total / 45.9B-active hybrid-linear-attention reasoning model with 1M context, 80K output) into a much smaller-active, coding-and-agentic-focused product — a notable scaling-down of MiniMax’s open-weights flagship, mirroring the broader open-MoE shift toward 10B-class activations that this concept page has been tracking. Filing this without a tech report leaves the moe-routing-design comparison (expert count, top-k, load-balancing) and the post-training recipe (closed-loop SWE synthesis? executable verifiers à la Qwen3-Coder-Next?) open.