---
license: mit
library_name: transformers
base_model: MiniMaxAI/Minimax-M3-preview
pipeline_tag: text-generation
tags:
  - eagle3
  - speculative-decoding
  - draft-model
  - gqa
  - vllm
  - torchspec
  - minimax
---

## Model Overview

**Inferact/MiniMax-M3-EAGLE3-GQA** is a **grouped-query-attention (GQA)** EAGLE3 draft model for accelerating inference of [MiniMax-M3](https://huggingface.co/MiniMaxAI/MiniMax-M3), served with **[vLLM](https://github.com/vllm-project/vllm)** and trained with **[TorchSpec](https://github.com/lightseekorg/TorchSpec)**.

It is **retrained on the same datasets** as the multi-head-attention version [Inferact/MiniMax-M3-EAGLE3](https://huggingface.co/Inferact/MiniMax-M3-EAGLE3) — **kimi-mtp, OpenCodeInstruct, SWE-bench, and SWE-bench-Pro** — with the draft's attention changed from **MHA to GQA** (`num_key_value_heads: 64 → 4`) for **inference efficiency** (16× smaller draft KV cache) and **compatibility with the target model**.

The draft is a 1-layer dense Llama (`LlamaForCausalLMEagle3`) on MiniMax-M3's `hidden_size=6144` / `vocab_size=200064`; at serve time it shares the target's embedding and LM head (EAGLE3). See `config.json` for the full architecture.

---

## Performance

Mean accepted length and draft accept rate measured end-to-end against `MiniMaxAI/MiniMax-M3-MXFP8` served with vLLM at `tensor-parallel-size=4`, `num_speculative_tokens=3`, greedy sampling (`temperature=0`, `top_p=1.0`), `max-concurrency=16`.

| Dataset | n | Mean accepted length | Draft accept rate | Per-position accept rate (pos 1 / 2 / 3) |
|---|---:|---:|---:|---:|
| MT-Bench | 64 | 2.668 | 55.62% | 0.745 / 0.537 / 0.387 |
| SPEED-Bench (qualitative) | 64 | 2.561 | 52.04% | 0.719 / 0.500 / 0.342 |