A 1-layer EAGLE3 speculative-decoding draft head for MiniMax-M2.7. It predicts draft tokens that the full M2.7 target model verifies, accelerating decoding without changing the target's outputs.

This is the full-vocabulary (200,064) draft — the highest-accuracy variant. For a smaller, faster-serving lm_head, see the vocab-pruned 32k variant.

Performance (mean speculative accept length)

Deployment-realistic config num_steps=3, topk=1, draft_tokens=4, bf16 draft:

benchmark	mean accept length
HumanEval	2.82
code agent (SWE-bench-Pro style)	2.85

Architecture

LlamaForCausalLMEagle3, num_hidden_layers=1, hidden_size=3072, vocab_size=200064, bf16.

Usage (SGLang)

bash

python -m sglang.launch_server \
  --model-path MiniMaxAI/MiniMax-M2.7 --tp 4 --trust-remote-code \
  --reasoning-parser minimax --tool-call-parser minimax-m2 \
  --attention-backend triton --speculative-draft-attention-backend triton \
  --speculative-algorithm EAGLE3 \
  --speculative-draft-model-path asherszhang/MiniMax-M2.7-EAGLE3-draft-vocab200k \
  --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 \
  --speculative-draft-model-quantization unquant

Notes: serve the draft as bfloat16 (M2.7 hidden states are bf16); pass unquant so the bf16 draft isn't force-quantized to the target's FP8.

License

Draft head weights released under apache-2.0. The model embeds embed_tokens from MiniMax-M2.7 and is a derivative — see the base model for its terms.