MiniMax-M3.1-Flash (428B MoE) Architecture Parameters, Benchmarks & VRAM
428B total / 23B active native multimodal 1M-context MoE with MiniMax Sparse Attention & DSpark draft head.
🧬 MiniMax-M3.1-Flash (428B MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (MiniMaxAI/MiniMax-M3.1-Flash) config.json & Model Card
1. Backbone & Dimensions
minimax_m3_vl- Total / Active Params
- 428B / 23B
- Hidden Size (d_model)
- 6,144
- FFN Intermediate Size
- 3,072
- Activation / Norm
- SiLU (SwiGLU) · ε=1e-6
- Native Tensor Dtype
- FP8 QAT
2. Attention & Layer Layout
HYBRID- Num Hidden Layers
- 60 Layers
- Hidden Layer Layout
- 60-Layer MiniMax Sparse Attention (MSA) + 128-Expert MoE (Top-4 + 1 Shared)
- Attention Heads (Q / KV)
- 64 Q / 4 KV (16:1 GQA)
- Head Dimension
- 128
- Linear / KV Compression
- End-to-end MiniMax Sparse Attention (MSA) + DSpark speculative head · 32L ViT
- Vision Encoder / MTP
- ViT 32L
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 1,048,576 tokens
- RoPE Configuration
- MSA Long-Context RoPE (原生 1,048,576 / 1M Context)
- Vocab Size
- 200,064 (Untied)
- Special Token IDs
- BOS: 200019 · EOS: 200020
📋 MiniMax-M3.1-Flash (428B MoE) Official API & System Card Specifications
Verified against official MiniMax developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- MiniMax-M3.1-Flash
- Release Date
- 2026-10-19
- Weight License
- Open-Weight (API Spec)
- Native Reasoning
- ✅ Supported
- Function Calling & JSON Schema
- 部分支持
- Batch API 异步批处理
- ✅ Supported (50% off)
2. Context, Output & Modalities
Limits- Max Context Window
- 1,024,000 tokens (1024K)
- Max Output Tokens
- 524,000 tokens
- Input Modalities
- Text + Image + Video
- Output Modalities
- Text
- Direct China Access
- ✅ Direct CN + RMB
- Deployment Mode
- Open-Weight + Cloud API
3. Official Pricing & Cache
CNY / 1M- Input / 1M Tokens
- ¥ 1
- Cached Input / 1M
- ¥ 0.1 -90%
- Output / 1M Tokens
- ¥ 4
- Tiered Pricing Notes
- MiniMax: Official MiniMax CN API endpoint
🏆 Official Benchmark Results
Sourced from official System Card / Model Card & technical reports
CyberGym is a benchmark for evaluating AI agents on cybersecurity tasks, testing their ability to identify vulnerabilities, perform security analysis, and complete security-related challenges in a controlled environment.
BankerToolBench is a public benchmark that evaluates models on banking and finance tool-use tasks. Models are scored against dataset rubrics, measuring their ability to correctly invoke tools and complete multi-step financial workflows.
OSWorld 2.0 is a benchmark of 108 long-horizon, real-world computer-use workflows spanning everyday and professional tasks.
Scientific chart synthesis and quantitative reasoning
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 MiniMax-M3.1-Flash (428B MoE) Full Benchmark Scorecard
13 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 5 items
CyberGym is a benchmark for evaluating AI agents on cybersecurity tasks, testing their ability to identify vulnerabilities, perform security analysis, and complete security-related challenges in a controlled environment.
DeepSWE 1.1 evaluates software engineering agents using the mini-swe-agent harness.
NL2Repo evaluates long-horizon coding capabilities including repository-level understanding, where models must generate or modify code across entire repositories from natural language specifications.
PostTrainBench evaluates a model's ability to autonomously post-train base models. Given pretrain-only base models, the agent must complete the full pipeline of data synthesis, training, evaluation, and iteration within a time budget, scored across downstream benchmarks such as…
Terminal-Bench 4.0 evaluates AI agents on real-world work in terminal and command-line containerized environments across 66 tasks, with emphasis on science-adjacent and frontier engineering problems.
🤖 Agents & Tools 4 items
BankerToolBench is a public benchmark that evaluates models on banking and finance tool-use tasks. Models are scored against dataset rubrics, measuring their ability to correctly invoke tools and complete multi-step financial workflows.
AutomationBench v1.0.6 evaluates agent performance on multi-step workflow automation tasks.
OSWorld 2.0 is a benchmark of 108 long-horizon, real-world computer-use workflows spanning everyday and professional tasks.
Agents' Last Exam is a challenging benchmark for AI agents on hard, long-horizon tasks that test sustained reasoning, planning, and tool use, reported with and without tool access.
📐 Math 1 items
👁️ Multimodal 3 items
A benchmark for early-stage visual reasoning and perception on child-like vision tasks.
CharXiv-R is the reasoning component of the CharXiv benchmark, focusing on complex reasoning questions that require synthesizing information across visual chart elements.
GDP.pdf is a knowledge-work vision benchmark that evaluates models on economically valuable professional tasks presented as visual documents (PDFs), testing document-based reasoning, chart and table interpretation, and problem solving without tools.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ MiniMax-M3.1-Flash (428B MoE) Interactive VRAM & Token Speed Calculator
Physical 512 GB · Load 58.2%
MoE: reads only 13.94 GB active weights/step
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run minimax-m3.1:flashMac Studio M5 Ultra 512G Fits MiniMax-M3.1-Flash (428B MoE) · Cost & TCO
Approx. $9,499 (~$19/GB VRAM) · 190W TDP
📊 MiniMax-M3.1-Flash (428B MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 856 GB | 2.01 GB | 876.43 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 454.75 GB | 2.01 GB | 467.16 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 350.96 GB | 2.01 GB | 361.29 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 304.42 GB | 2.01 GB | 313.82 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 259.47 GB | 2.01 GB | 267.98 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 227.38 GB | 2.01 GB | 235.24 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 209.19 GB | 2.01 GB | 216.68 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 168.53 GB | 2.01 GB | 175.21 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for MiniMax-M3.1-Flash (428B MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of MiniMax-M3.1-Flash (428B MoE)?
MiniMax-M3.1-Flash (428B MoE) uses MiniMaxM3SparseForConditionalGeneration with 60 layers (60-Layer MiniMax Sparse Attention (MSA) + 128-Expert MoE (Top-4 + 1 Shared)), hidden_size=6144, 64 Q heads / 4 KV heads (head_dim=128), and vocab_size=200,064.
How much VRAM does MiniMax-M3.1-Flash (428B MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), MiniMax-M3.1-Flash (428B MoE) requires 267.98 GB VRAM. Full FP16 requires 876.43 GB.