Skip to main content
Back

MiniMax-M3 Runs a 428B Frontier Model on a $34K Workstation Rig

A patched vLLM Docker image lets you self-host MiniMax-M3, a 428B frontier MoE, on four consumer Blackwell GPUs with 250K context — fixing three bugs upstream vLLM still ships without.

  • What: Community repo ships a patched vLLM Docker image to run MiniMax-M3 (428B MoE) on 4x RTX PRO 6000 Blackwell GPUs with 250K context.
  • Why it matters: Upstream vLLM silently degrades M3 outputs due to 3 missing patches; this repo fixes all of them in one image.
  • Performance: ~113 tok/s single-stream decode, ~184 tok/s at 4 concurrent requests, ~2,800 tok/s prefill up to 131K context.
  • MiniMax-M3: 59.0% SWE-Bench Pro, 1M-token context via MSA sparse attention (15.6x faster decode than M2 at 1M tokens), native multimodal.
  • Hardware: 4x RTX PRO 6000 = 384GB GDDR7 VRAM (~$34K), no NVLink — PCIe Gen 5 is the bottleneck for multi-GPU scaling.
  • Limitation: vLLM path has no vision tower support; use SGLang path for image/video inference.

MiniMax-M3 is a frontier-class open-weight model that most people assumed required a rack of H100s to self-host. A new community repo called minimax-m3-sm120 changes that math. It ships a patched vLLM Docker image and a complete launch recipe for running M3 on four RTX PRO 6000 Blackwell GPUs , a workstation-class setup that costs a fraction of datacenter hardware , with a 250K context window and verified throughput numbers.

What you're actually running

MiniMax-M3 is a 428B / 23B-active Mixture-of-Experts model with MiniMax Sparse Attention (MSA), and this repo serves the MXFP4-quantized checkpoint at up to 250K context on consumer Blackwell GPUs. The model itself is a significant release: MiniMax M3 launched as the first open-weight model to simultaneously hit a 59.0% SWE-Bench Pro score, support a 1M-token context window, and ship native image and video understanding in a single checkpoint.

The architecture's headline innovation is MSA. MSA factors attention into two stages: an Index Branch and a Main Branch. The Index Branch decides which key-value blocks each query should read. The Main Branch then runs exact softmax attention over only those selected blocks. The default block size is 128 tokens. Each query and GQA group keeps 16 blocks, fixing the per-query budget at 2,048 key-value tokens , regardless of how long the full context is. The payoff is dramatic: MiniMax reports per-token compute at 1M context reduced to roughly 1/20th of the previous generation, with 9x prefill and over 15x decode speedup versus dense baselines.

Unlike DeepSeek's Multi-head Latent Attention (MLA), which compresses keys and values into a low-dimensional latent space, MSA operates on a standard GQA backbone but uses block-level selection on real, uncompressed key-values. This is a deliberate design choice: M3 keeps a GQA backbone but layers MSA on top, doing block-level selection on real, uncompressed key-values , deliberately different from MLA, which compresses KV state and trades off some long-context precision.