MiniMax-M3 (427B MoE) Architecture Parameters, Benchmarks & VRAM
MiniMax M3 427B/A23B multimodal sparse-attention MoE with 1M context and 524K max output.
🧬 MiniMax-M3 (427B MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (MiniMaxAI/MiniMax-M3) config.json & Model Card
1. Backbone & Dimensions
minimax_m3_vl- Total / Active Params
- 427B / 23B
- Hidden Size (d_model)
- 6,144
- FFN Intermediate Size
- 3,072
- MoE Experts
- 128 Routed · Top-4 + 1 Shared
- Activation / Norm
- SiLU (SwiGLU) · ε=0.000001
- Native Tensor Dtype
- BF16 / FP8
2. Attention & Layer Layout
GQA- Num Hidden Layers
- 60 Layers
- Hidden Layer Layout
- 60-Layer Sparse MoE (128 Experts, Top-4 Active)
- Attention Heads (Q / KV)
- 64 Q / 4 KV (16:1 GQA)
- Head Dimension
- 128
- Vision Encoder / MTP
- ViT 32L
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 1,048,576 tokens
- RoPE Configuration
- RoPE (原生 1,048,576 Context)
- Vocab Size
- 200,064 (Untied)
📋 MiniMax-M3 (427B MoE) Official API & System Card Specifications
Verified against official MiniMax developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- MiniMax-M3
- Release Date
- 2026-06-01
- 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
- 1,000,000 tokens (1000K)
- 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
- ¥ 2.1
- Cached Input / 1M
- ¥ 0.42 -80%
- Output / 1M Tokens
- ¥ 8.5
- Tiered Pricing Notes
- MiniMax: Official MiniMax CN API endpoint
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 MiniMax-M3 (427B MoE) Full Benchmark Scorecard
25 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 11 items
A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.
Claw-Eval tests real-world agentic task completion across complex multi-step scenarios, evaluating a model's ability to use tools, navigate environments, and complete end-to-end tasks autonomously.
MCP Atlas is a benchmark for evaluating AI models on scaled tool use capabilities, measuring how well models can coordinate and utilize multiple tools across complex multi-step tasks.
Terminal-Bench 2.1 is an updated release of the Terminal-Bench benchmark that tests AI agents' ability to operate a computer via the terminal.
SWE-Bench Pro is an advanced version of SWE-Bench that evaluates language models on complex, real-world software engineering tasks requiring extended reasoning and multi-step problem solving.
PaperBench is a benchmark for evaluating AI agents on their ability to replicate research papers. It tests models on complex, multi-step workflows involving code implementation, experimentation, and reproducing scientific results from academic publications.
NL2Repo evaluates long-horizon coding capabilities including repository-level understanding, where models must generate or modify code across entire repositories from natural language specifications.
SWE Atlas - Codebase QnA evaluates a model's ability to answer questions about real codebases, measuring repository-level comprehension and the ability to reason about code structure, behavior, and intent across an entire project.
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…
CL-bench is an open-source benchmark with its own data and rubrics for evaluating models on coding and agentic tasks, scored using a setup fully aligned with the official procedure.
FrontierCode 1.1 evaluates whether coding-agent changes are mergeable, using unit tests, maintainer-defined rubrics, and verifiers. Runs flagged for unfair internet use receive a zero score.
🤖 Agents & Tools 8 items
SpreadSheetBench-v1 evaluates office automation agents on spreadsheet reasoning and manipulation tasks, measuring the ability to analyze, transform, and operate on spreadsheet data through tools.
BrowseComp is a benchmark comprising 1,266 questions that challenge AI agents to persistently navigate the internet in search of hard-to-find, entangled information.
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.
DRACO is a deep research benchmark that evaluates an agent's ability to gather, synthesize, and reason over information to answer complex research questions. Scores are based on official rubrics per question, with the final score being the average across all questions.
OSWorld-Verified is a verified subset of OSWorld, a scalable real computer environment for multimodal agents supporting task setup, execution-based evaluation, and interactive learning across Ubuntu, Windows, and macOS.
Finance Agent v2 is an agentic financial-analysis benchmark from Vals that evaluates models on real-world finance workflows, measuring their ability to retrieve and reason over financial documents, perform multi-step calculations, and produce accurate analyses.
OfficeQA Pro evaluates AI models on professional knowledge-work questions and tasks drawn from real office workflows, including document analysis, spreadsheet reasoning, and information synthesis across business domains.
APEX-Agents is a benchmark evaluating AI agents on long horizon professional tasks that require sustained reasoning, planning, and execution across complex multi-step workflows.
📐 Math 2 items
USAMO 2026 evaluates models on the six problems from the 2026 United States of America Mathematical Olympiad, requiring rigorous proof-based reasoning.
LiveBench is a challenging, contamination-limited LLM benchmark that addresses test set contamination by releasing new questions monthly based on recently-released datasets, arXiv papers, news articles, and IMDb movie synopses.
👁️ Multimodal 4 items
OmniDocBench 1.5 is a comprehensive benchmark for evaluating multimodal large language models on document understanding tasks, including OCR, document parsing, information extraction, and visual question answering across diverse document types.
Video-MME is the first-ever comprehensive evaluation benchmark of Multi-modal Large Language Models (MLLMs) in video analysis.
Video-MMMU evaluates Large Multimodal Models' ability to acquire knowledge from expert-level professional videos across six disciplines through three cognitive stages: perception, comprehension, and adaptation.
A more robust multi-discipline multimodal understanding benchmark that enhances MMMU through a three-step process: filtering text-only answerable questions, augmenting candidate options, and introducing vision-only input settings.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ MiniMax-M3 (427B MoE) Interactive VRAM & Token Speed Calculator
Physical 512 GB · Load 58%
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-m3-428bMac Studio M5 Ultra 512G Fits MiniMax-M3 (427B MoE) · Cost & TCO
Approx. $9,499 (~$19/GB VRAM) · 190W TDP
📊 MiniMax-M3 (427B MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 854 GB | 2.01 GB | 874.39 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 453.69 GB | 2.01 GB | 466.07 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 350.14 GB | 2.01 GB | 360.46 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 303.7 GB | 2.01 GB | 313.09 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 258.87 GB | 2.01 GB | 267.36 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 226.84 GB | 2.01 GB | 234.69 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 208.7 GB | 2.01 GB | 216.18 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 168.13 GB | 2.01 GB | 174.81 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for MiniMax-M3 (427B MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of MiniMax-M3 (427B MoE)?
MiniMax-M3 (427B MoE) uses MiniMaxM3SparseForConditionalGeneration with 60 layers (60-Layer Sparse MoE (128 Experts, Top-4 Active)), hidden_size=6144, 64 Q heads / 4 KV heads (head_dim=128), and vocab_size=200,064.
How much VRAM does MiniMax-M3 (427B MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), MiniMax-M3 (427B MoE) requires 267.36 GB VRAM. Full FP16 requires 874.39 GB.