MiniMax-M3-EAGLE3.1

Model Overview

• Model Architecture: LlamaForCausalLMEagle3 (EAGLE3 speculative-decoding draft model)
  • Input: Text
  • Output: Text
• Draft variant: EAGLE3.1 (single decoder layer, full vocabulary, BF16)
• Target model: amd/MiniMax-M3-MXFP4
• Supported Hardware Microarchitecture: AMD Instinct MI350X / MI355X
• Inference Engine: vLLM (ROCm)
• Trained by: the AMD Quark team

MiniMax-M3-EAGLE3.1 is an EAGLE3 draft model for accelerating inference of amd/MiniMax-M3-MXFP4 with speculative decoding. It was trained from scratch (cold-start) entirely on AMD Instinct MI350X GPUs with a vLLM-centric pipeline (on-policy data generated by the target, hidden-state extraction from the target, and in-loop serve-evaluation). Speculative decoding is lossless — every draft token is verified by the target, so the target's output distribution is preserved.

Intended Use

This model is intended to be used as an EAGLE3 draft model for speculative decoding with amd/MiniMax-M3-MXFP4 as the target model. It reuses the target's tokenizer, so no tokenizer files are shipped with the draft.

Acceptance Length

Evaluated on the official SPEED-Bench harness (specdec_bench) at TP=8 on vLLM ROCm v0.27.1 with native AITER MXFP4 MoE kernels, using num_speculative_tokens=3 and temperature 0. Qualitative results are the mean of two complete runs; each fixed-context Throughput split is one complete run. Acceptance length (AL) is the mean number of tokens emitted per target-model verification step (AL = 1 means no speedup); higher is better.

Acceptance length by domain (SPEED-Bench qualitative)

Acceptance length by context length

Acceptance length stays essentially flat at 2.71–2.72 from 1K to 32K context.

Native AITER Serving Results

On SPEED-Bench Qualitative at concurrency 48, EAGLE3 improves output throughput and end-to-end latency over a matched non-speculative baseline:

The EAGLE3 values are means of two complete runs. Both configurations use vLLM ROCm v0.27.1, native AITER MXFP4, TP=8, temperature 0, a 12,288-token maximum model length, and a 2,048-token output cap. The TTFT result shows the first-token latency trade-off.

Across concurrency 1–128, EAGLE3 improves per-GPU output throughput at every tested point, scaling from 29 to 652 tok/s/GPU versus 16 to 465 tok/s/GPU for the non-speculative baseline. The relative gain narrows from 1.77× at concurrency 1 to 1.40× at concurrency 128 as the baseline approaches saturation, while the absolute throughput advantage grows from approximately 13 to 187 tok/s/GPU. Figure labels are rounded; speedup ratios use the underlying unrounded benchmark values.

Serving with vLLM

Serve the amd/MiniMax-M3-MXFP4 target with this draft as the EAGLE3 speculative model at TP=8. See the target model card for the ROCm/vLLM runtime image and setup.

export VLLM_ROCM_USE_AITER=1
vllm serve amd/MiniMax-M3-MXFP4 --trust-remote-code --tensor-parallel-size 8 \
    --block-size 128 --attention-backend TRITON_ATTN --moe-backend aiter \
    --speculative-config '{"method":"eagle3","model":"amd/MiniMax-M3-EAGLE3.1","num_speculative_tokens":3,"attention_backend":"TRITON_ATTN"}'

Citation and Acknowledgements

Trained by the AMD Quark team as the EAGLE3 draft for amd/MiniMax-M3-MXFP4. Please validate quality and acceptance length in your own serving stack.

License

This draft targets MiniMax-M3; see the bundled MiniMax M3 LICENSE.txt (MiniMax Community License) for the terms that apply to the target model and its derivatives.

Modifications Copyright (c) 2026 Advanced Micro Devices, Inc. All rights reserved.