MiniMax M3
MiniMax M3 is a 428B-parameter model you reach through an API, with 23B parameters active per token — too large to self-host, so this page covers what it does, how to access it, and what to run locally instead.
You don't run MiniMax M3 locally
At 428B parameters (23B active per token), a Q4-class build of MiniMax M3 would need roughly 257 GB — only a maxed-out 512GB Mac Studio could even hold the weights, with no real headroom (0.6 GB per billion parameters, our standard Q4 rule). Access is through the hosted API.
Jun 1, 2026 MiniMax release. Native-multimodal MoE (~428B total / ~23B active) with 1M context via MiniMax Sparse Attention. SWE-Bench Pro 59.0%, Terminal-Bench 2.1 66.0%. Open weights; cloud/API for nearly all users.
Strong local alternatives
Frequently asked questions
Can I run MiniMax M3 locally?
Not realistically. MiniMax M3 is a 428B-parameter model (23B active); a Q4-class build would need roughly 257 GB — only a maxed-out 512GB Mac Studio could even hold the weights, with no real headroom. The hosted API or a smaller open model is the practical path.
How do I access MiniMax M3?
Through the vendor-hosted API. See the official source linked on this page.
What is the best local alternative to MiniMax M3?
Qwen3 235B A22B is the strongest local model we track (235B). It runs on a single high-memory Mac or GPU; see its page for exact hardware.
Cite this page
ModelFit: MiniMax M3 — specs, memory math and hardware verdicts. https://modelfit.io/models/minimax-m3/ (dataset updated 2026-09-03, CC BY 4.0).