MiniMax M3.1 Flash open weights: 428B parameters, downloads and version differences

Four entries, different evidence

MiniMax's identification of Space Bunny establishes product lineage. Dates, routes, reasoning settings and serving implementations differ, so these records do not form a controlled before/after comparison. Weight availability alone adds no new TOKRACE score.

What is in the public model

Architecture

Native multimodal mixture-of-experts (MoE)

Inputs → output

Text, images and video → text

Context

Up to 1,000,000 tokens; usable length depends on hardware and serving configuration

Attention

MiniMax Sparse Attention (MSA) throughout, including the first three layers

Training / serving

Quantization-aware training; DSpark draft head for compatible serving implementations

Reasoning settings

Chat-template reasoning_effort: low, medium, high, xhigh, max

The model card describes three changes from M3: sparse attention replaces the first three full-attention layers, quantization-aware training reduces the serving footprint, and DSpark replaces the earlier EAGLE-like MTP approach. A 23B active-parameter count describes computation per token; all experts still need to be stored or loaded. It is not a 23B model download or a memory requirement. Official model card

Official benchmarks: reported for the official weights

The following five results are published by MiniMax. TOKRACE has not independently reproduced them. They remain separate from the historical API task tests.

Agent benchmarks depend on the harness, context and tool environment as well as the model. The model card also cautions that third-party quantizations may differ from the official weights. Read the complete official table and notes

Download weights or use a hosted service

Hugging Face · ModelScope

The Hugging Face repository was public and ungated when checked. Its file listing contains 48 main weight shards and a separate DSpark draft head. We checked repository metadata and a weight-file response; we did not download the full model or validate local serving.

The listed weight files total approximately 252.4 GB (decimal file sizes, not a runtime-memory estimate).

With the Hugging Face CLI installed, pin the checked revision for a reproducible download:

hf download MiniMaxAI/MiniMax-M3.1-Flash \
  --revision 54e4e69dfc452b1e94a426c24a4fdeeaa8e3ad5e \
  --local-dir MiniMax-M3.1-Flash

The model card links serving recipes for SGLang · vLLM · TokenSpeed. Check each recipe's hardware, quantization and DSpark support.

Hosted access remains separate: MiniMax API · M Plan · MiniMax Code desktop. Their availability and billing do not describe local-weight inference costs.

MiniMax Community License

The repository uses the MiniMax Community License. Its commercial-use terms require visible “Built with MiniMax M3.1-Flash” attribution. The license text requires prior written authorization for products and services generating more than US$20 million in yearly revenue; otherwise it calls for a one-time notice. It also lists prohibited uses.

Read the license at the checked revision · Current repository license

Sources and verification record

Official specifications and scores come from MiniMax. TOKRACE's additions are the source checks and the explanation of how the release differs from earlier API evidence.

• Initial repository commit · Checked snapshot
• Official English model card · Checked snapshot
• Official Chinese model card · Checked snapshot
• MiniMax Community License · Checked snapshot
• October 19 API launch and Space Bunny identity announcement

Download verification manifest, source hashes and revision (JSON)

Checked revision: 54e4e69dfc452b1e94a426c24a4fdeeaa8e3ad5e