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

MiniMax Releases 428B Multimodal MoE Model MiniMax-M3 With 1M Context and Sparse Attention

MiniMax has published weights for MiniMax-M3, a 428B-parameter multimodal Mixture-of-Experts model with 23B active parameters, a one-million-token context window, and MiniMax Sparse Attention.

Listen to this articleBrowser Voice
Ready
AI-generated conceptual illustration for MiniMax Releases 428B Multimodal MoE Model MiniMax-M3 With 1M Context and Sparse Attention; 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.

MiniMax has released the weights and deployment tooling for MiniMax-M3, a native multimodal Mixture-of-Experts (MoE) model supporting a one-million-token context window across text, image, and video inputs. Published on the MiniMaxAI/MiniMax-M3 Hugging Face repository, the release pairs roughly 428 billion total parameters with an active footprint of about 23 billion parameters per token. Built for long-horizon coding, agentic workflows, and multimodal reasoning, the model introduces MiniMax Sparse Attention (MSA) alongside configurable thinking modes and broad serving support.

Native multimodal scope and ecosystem positioning

Rather than attaching vision adapters onto a text-only foundation model late in training, MiniMax states that M3 undergoes mixed-modality training from the very first step to enable deeper semantic fusion across text, images, and video. Categorized under minimax_m3_vl and image-text-to-text, the model targets conversational assistants, repository-scale coding, and multi-step cowork automation. On Hugging Face, the release has gained rapid developer traction, recording 104,156 monthly downloads, 56 community quantizations, 15 fine-tunes, and 70 Spaces, including agentic workflows such as smolagents/ml-intern and FINAL-Bench/ViKernel-Scanner. Notably, the model card already points visitors to a newer variant, MiniMaxAI/MiniMax-M3.1-Flash, reflecting fast iteration within the M3 lineup.

Mixture-of-Experts scale and MiniMax Sparse Attention

Safetensors metadata on the repository lists 427 billion total parameters in BF16 and F32 tensor formats, while MiniMax's overview cites approximately 428 billion total parameters with 23 billion activated per token via MoE routing. The central architectural update is MiniMax Sparse Attention (MSA), described in an accompanying June technical report (arXiv:2606.13392). According to MiniMax, MSA is a sparse attention operator engineered for million-token contexts that cuts attention compute and memory overhead compared with Grouped-Query Attention (GQA) while preserving output quality. Against its predecessor M2 at a one-million-token context length, MiniMax claims 9x faster prefill, 15x faster decode throughput, and one-twentieth of the per-token compute. Those efficiency gains are vendor-reported figures from the model card and paper rather than results verified by independent third-party testing.

Evaluation claims across coding, video, and agentic tasks

The Hugging Face card highlights reported scores across several public coding, multimodal, and terminal-agent leaderboards. In software engineering, MiniMax-M3 records 80.5 on SWE-bench Verified and 59 on ScaleAI's SWE-bench Pro. For multimodal and video tasks, the card lists 78.1 on MMMU Pro (Standard 10 Options) and 85.4 on Video-MME-v2. On IntelligenceLab's Long-Horizon-Terminal-Bench using the official LHTB Harbor harness, the model logs a mean reward of 38.5 (multiplied by 100) across 46 tasks, solving 3 at the 0.95 pass threshold. Under the expanded LHTB 2.0 benchmark of 77 tasks with a three-hour per-task budget on terminus-2, it posts a mean reward of 46.9 and solves 5 tasks at 0.95 or higher—illustrating both capability gains and the high failure rate that persists in autonomous multi-hour terminal execution.

Deployment stack, reasoning controls, and developer workflows

MiniMax-M3 arrives with documented configurations across major inference engines and OpenAI-compatible APIs. Developers can load the model in Hugging Face transformers using AutoProcessor and AutoModelForMultimodalLM, launch servers with vllm serve or SGLang's sglang.launch_server and lmsysorg/sglang:latest Docker container, or run it via Docker Model Runner (docker model run hf.co/MiniMaxAI/MiniMax-M3). Additional instructions cover KTransformers, Unsloth, and ATOM for MXFP4 and MXFP8 deployment, alongside hosted access through the MiniMax API and four Hugging Face inference providers including OpenRouter. Through the thinking parameter, applications can switch between enabled, adaptive, and disabled reasoning modes to trade latency for deeper deliberation, with MiniMax recommending temperature=1.0 and top_p=0.95.

Licensing, hardware demands, and operational caveats

Although weights are publicly downloadable, MiniMax-M3 is not distributed under an OSI-approved open-source license. Instead, it is governed by the custom minimax-community license, requiring legal teams to inspect specific terms covering commercial use, redistribution, and derivative works before production adoption. Hardware requirements also remain substantial: serving 427B-plus parameters in BF16 or F32 demands significant multi-GPU VRAM capacity despite the 23B active parameter count, pushing smaller teams toward MXFP4/MXFP8 guides, community quantizations for llama.cpp or Ollama, or managed API endpoints. Finally, the repository uses custom_code and requires trust_remote_code=True in Transformers, meaning engineering teams should audit remote modeling scripts and isolate execution environments before deployment.

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-19.

Follow your curiosity.

Pass the signal.