Skip to content
AIBRAINWORX / THE INDEPENDENT AI LABRESEARCH · HARDWARE · OPEN INTELLIGENCE
Models & Weights / Inference & Hardware / Model Watch

MiniMax Releases MiniMax-M2, a 230B Sparse MoE Model Built for Coding and Agentic Workflows

MiniMax has published MiniMax-M2 on Hugging Face, pairing 230 billion total parameters with 10 billion active parameters per token for coding loops, terminal tools, and web agents.

Listen to this articleBrowser Voice
Ready
AI-generated conceptual illustration for MiniMax Releases MiniMax-M2, a 230B Sparse MoE Model Built for Coding and Agentic Workflows; not a photograph or product screenshot
AI-generated conceptual illustration by AiBrainWorX; not a photograph or product screenshot.
Event / context: 2026-11-19. Source check: October 28, 2026.

Sparse MoE Release Targets Coding and Agentic Toolchains

MiniMax has published MiniMaxAI/MiniMax-M2 on Hugging Face, introducing an open-weight Mixture-of-Experts (MoE) language model engineered specifically for software engineering loops, terminal execution, and multi-step agentic tool use. Listed on November 19, 2026 under the minimax_m2 architecture, the release pairs a 230-billion-parameter total footprint with 10 billion activated parameters per token. MiniMax positions the model as a compact-compute alternative for interactive developer agents and batched sampling, rather than a monolithic dense model requiring full-parameter activation on every forward pass. Alongside the Hugging Face weights in Safetensors format, MiniMax points developers to ModelScope, GitHub, Model Context Protocol (MCP) integrations, MiniMax Code, and a hosted API offered free for a limited time.

Architecture, Inference Specs, and Reported Benchmarks

Technically, MiniMax-M2 relies on a sparse MoE design with 230 billion total parameters and 10 billion active parameters, tagged on Hugging Face with fp8, safetensors, transformers, and custom_code. Because it uses custom modeling code, direct loading via Hugging Face transformers requires passing trust_remote_code=True to AutoTokenizer, AutoModelForCausalLM, or the text-generation pipeline. The model card also references three arXiv papers (2504.07164, 2509.06501, and 2509.13160) and highlights support across major inference engines including vLLM, SGLang, MLX, and Docker Model Runner.

On performance, MiniMax attributes a #1 global open-source composite intelligence ranking to benchmarks compiled by Artificial Analysis, spanning mathematics, science, instruction following, coding, and tool use. In its own model-card evaluation suite, MiniMax reports results across coding and web-agent benchmarks compared against Claude Sonnet 4, Claude Sonnet 4.5, Gemini 2.5 Pro, GPT-5 (thinking), GLM-4.6, Kimi K2 0905, and DeepSeek-V3.2:

  • Software engineering and terminal suites: MiniMax-M2 scores 69.4 on SWE-bench Verified (run on internal infrastructure using an R2E-Gym scaffold atop OpenHands at 128k context length, 100 maximum steps, and no test-time scaling), 36.2 on Multi-SWE-Bench, 56.5 on SWE-bench Multilingual, and 46.3 on Terminal-Bench using the official claude-code CLI averaged across eight runs.
  • Web search and deep-research agents: Using the WebExplorer agent framework, MiniMax reports 44 on BrowseComp, 48.5 on BrowseComp-zh, 75.7 on the 103-sample text-only GAIA validation subset, 72 on xbench-DeepSearch, and 31.8 on the text-only Humanity's Last Exam (HLE) subset with search and Jupyter Python tools.
  • Tool-use and domain tasks: The card lists 66.8 on ArtifactsBench (judged by Gemini-2.5-Pro), 77.2 on τ²-Bench (using extended thinking with tool use and a GPT-4.1 user simulator), 65.5 on FinSearchComp-global, and 36 on AgentCompany.

All benchmark figures are self-reported by MiniMax or drawn from third-party model reports noted in the table, rather than verified through independent testing by this publication.

Deployment Workflows for Local and Server Teams

For engineering teams building coding assistants, CI repair pipelines, or browser-equipped agents, the primary practical impact lies in decoupling active inference cost from total parameter count. Activating 10 billion parameters per generation step is designed to reduce per-token latency and improve throughput for multi-turn agent loops, where a single bug fix or search workflow can require dozens of sequential shell commands, file edits, and test runs.

Out-of-the-box serving recipes lower integration friction across production and workstation environments. Teams running OpenAI-compatible endpoints can serve the model directly via vllm serve "MiniMaxAI/MiniMax-M2" or launch an SGLang server with python3 -m sglang.launch_server on port 30000. For containerized GPU deployments, the card documents the lmsysorg/sglang:latest Docker image configured with --gpus all, --shm-size 32g, and --ipc=host. Individual developers are also directed to docker model run hf.co/MiniMaxAI/MiniMax-M2, MLX for Apple silicon workflows, and community quantizations compatible with llama.cpp, Ollama, and LM Studio.

Hardware Demands, Licensing Ambiguity, and Evaluation Caveats

Despite its 10-billion active parameter profile, MiniMax-M2 is not a lightweight download in terms of memory footprint. Hosting a 230-billion-parameter MoE model still requires substantial aggregate VRAM or unified memory, even in FP8 or quantized formats, meaning local workstation users will rely on heavy quantization while full-capacity serving remains geared toward multi-GPU server nodes. Requiring trust_remote_code=True in Transformers also introduces code-execution risk if organizations do not pin revisions and audit custom repository files before deployment.

Two additional caveats merit attention from compliance and procurement teams. First, the Hugging Face repository metadata lists the license as modified-mit, whereas the model card body states License: MIT; commercial adopters should inspect the exact license text in the repository rather than assuming uniform permissive terms. Second, the Hugging Face listing already notes a newer revision, MiniMaxAI/MiniMax-M2.7, signaling rapid iteration on the M2 family. Finally, because MiniMax's coding and agent scores depend on specific scaffolds—including OpenHands, claude-code, WebExplorer, and LLM judges like Gemini-2.5-Pro and GPT-4.1—real-world performance in production IDEs or enterprise repositories will vary with prompt design, harness engineering, and tool reliability.

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-20.

Follow your curiosity.

Pass the signal.