MiniMax M3: The Open-Weight 1M-Context Frontier Model

Quick verdict

MiniMax M3 matters because of what it bundles into a model you can download. Until now, a 1M-token context, native video understanding, and frontier-tier agentic coding tended to be three different products — usually closed ones. M3 puts them in one ~428B MoE and ships the weights. If you want a self-hostable model for long-document analysis, computer-use agents, or multimodal pipelines, it's the standout pick of mid-2026.

The honest caveat: the headline benchmark numbers are MiniMax's own, run on their infrastructure with baselines they chose. At least one outlet flagged that the figures have not been independently verified — Scale AI, which maintains the official SWE-Bench Pro benchmark, has not confirmed them. Treat the "beats GPT-5.5 / Gemini 3.1 Pro" framing as a vendor claim until third parties replicate it. For privacy-first, local-first work, M3 is still a genuine step forward over its predecessor, DeepSeek V4, and Kimi K2.6 (now succeeded by Kimi K3) on multimodal + long-context throughput.

Specs at a glance

MiniMax Sparse Attention (MSA): the real headline

Dense attention cost is what has historically made 1M-token context impractical to run locally — it scales badly as the context grows. MSA attacks that directly. A lightweight index branch scans incoming tokens and picks which blocks of past tokens actually need attention, then runs attention only on those blocks. The result, per MiniMax: at 1M-token context, per-token compute drops to roughly 1/20 of the previous generation.

In practical terms, MSA is what makes a 1M-context model viable to actually serve. The figures are MiniMax's, from its launch post and model card. The technical report, MiniMax Sparse Attention (arXiv:2606.13392), landed on June 11, 2026, so the mechanism is now documented rather than just announced.

Benchmarks (vendor-reported)

Read this section with one caveat in mind: every number below is MiniMax-reported, run on MiniMax's own infrastructure with baselines they selected. None have been independently confirmed by Scale AI (which maintains SWE-Bench Pro), and at least one outlet flagged the comparison as not independently verified. They are encouraging, not settled.

Sources: MiniMax M3 launch blog and technical report (June 2026); VentureBeat launch coverage. All scores are vendor-run; the SWE-Bench Pro comparison to GPT-5.5 and Gemini 3.1 Pro is MiniMax's claim and has not been independently verified. We attribute rather than assert it.

Native multimodality

M3 is natively multimodal — it ingests image and video input directly rather than relying on an upstream frame-extraction or OCR step, and it can operate a desktop computer for agentic workflows. Combined with the 1M-token context, that makes it a single model for tasks that previously needed a chain of specialized tools: read a long video, hold an entire codebase or document corpus in context, and act on it.

Pricing

If you use the hosted API (MiniMax or third-party gateways like OpenRouter), M3 is priced aggressively — a fraction of frontier closed models. Note the tiered structure: the standard rate applies to requests up to 512K input tokens; above that, the request is billed at double the rate. Prices below are from MiniMax's pricing page, read 2 November 2026; a Priority tier costs 1.5x these.

MiniMax labels the 50% discount permanent, and OpenRouter lists the same $0.30 / $1.20. Because M3 is open weight, self-hosting is the other lever, but at 257-265 GB for a 4-bit build it only pays off on hardware you already own or for sustained high-volume work.

Can you run MiniMax M3 locally?

Yes, the weights are downloadable, but not on ordinary hardware: even the smallest 1-bit GGUF is 129 GB, a 4-bit build is about 257-265 GB, and the official FP8 weights are 444 GB. And there is no local Ollama tag: minimax-m3:cloud runs on Ollama's servers. In practice, running M3 yourself means a Mac Studio with 256-512 GB of unified memory, a multi-GPU server, or a CPU-plus-GPU setup with several hundred gigabytes of system RAM.

Sizes below were read on 2 November 2026 from the MiniMaxAI/MiniMax-M3, MiniMax-M3-MXFP8 and unsloth/MiniMax-M3-GGUF repositories (file totals from the Hugging Face API), and the Ollama registry. Add memory for the KV cache and the vision encoder on top of the weights.

• •It is bigger than its launch coverage suggested. MiniMax's Hugging Face card gives ~428B total parameters with ~23B active per token, and the safetensors hold 427.0 billion. At BF16 that is 854 GB of weights alone.
• •Quantization is the path to a single box. A 4-bit GGUF is about 257-265 GB, which is the realistic target for a 512 GB Mac Studio or a server with four 80 GB cards. A 3-bit build (196 GB) is the floor for a 256 GB machine.
• •Only ~23B params activate per token. The MoE design (128 routed experts plus one shared, 4 routed per token) keeps compute per token far below the total, which is why CPU-plus-GPU runtimes that keep the experts in system RAM are viable. You still need memory to hold all of it. This page gives no tokens-per-second figure for an MoE model.
• •MSA helps the most at long context. The sparse-attention speedups are what make a long prompt tractable on your own hardware rather than purely theoretical.

MiniMax's card recommends SGLang or vLLM for serving, with Transformers, KTransformers (which keeps the experts in CPU memory) and unsloth's GGUFs as the other routes. To get the weights, or to try it through Ollama's cloud tier:

# Original weights, 854 GB (from MiniMax's model card)
hf download MiniMaxAI/MiniMax-M3 --local-dir MiniMax-M3

# Ollama: cloud only, nothing is downloaded to your machine
ollama run minimax-m3:cloud

New to quantization and VRAM math? Start with our quantization guide and the VRAM requirements guide. If your hardware can't hold a 428B model, a smaller open-weight coder from the best local AI models for programming roundup is the pragmatic choice, and the hosted API is a cheap stopgap while you size a rig. We will re-check the local options (including any MiniMax-M3.1-Flash builds) in the next refresh.

When to pick MiniMax M3 — and when not to

For a full side-by-side of the free models you can run, see our free local AI models guide roundup and the model comparisons hub.