nebius/EAGLE3-gpt-oss-120b

Model Description

This is an EAGLE-3 draft-model for gpt-oss-120b, trained from scratch using LK losses — training objectives that directly target acceptance rate rather than using KL divergence as a proxy.

Training Details

• Base model: openai/gpt-oss-120b
• Draft architecture: EAGLE-3
• Training data: Infinity-Instruct-0625 with gpt-oss-120b generated responses
• Training objective: Hybrid LK loss with adaptive λ scheduling (η=3)
• Training: 10 epochs from random initialization
• Draft length: K = 6 speculative tokens

Performance

Average acceptance length (τ) measured across MT-bench, HumanEval, and GSM8K with K = 7:

Configuration	Temperature = 0	Temperature = 1
EAGLE-3 + KL	2.76	2.46
EAGLE-3 + LK (ours)	2.81	2.65

Usage with vLLM

from vllm import LLM, SamplingParams

llm = LLM(
    model="openai/gpt-oss-120b",
    speculative_config={
        "method": "eagle3",
        "model": "nebius/EAGLE3-gpt-oss-120b",
        "num_speculative_tokens": 6,
    },
)

sampling_params = SamplingParams(temperature=0.7)
outputs = llm.generate(["Explain speculative decoding in simple terms."], sampling_params)

Note: The current vLLM implementation samples draft tokens greedily regardless of temperature settings, which can underestimate acceptance rates at temperature > 0. A community fix is under development (see vllm-project/vllm#20459). The acceptance metrics reported above were measured with proper rejection sampling.

License

CC BY 4.0

Citation

@misc{samarin2026lklosses,
  title     = {LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding},
  author    = {Alexander Samarin and Sergei Krutikov and Anton Shevtsov and Sergei Skvortsov and Filipp Fisin and Alexander Golubev},
  year      = {2026},
  eprint    = {2602.23881},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url       = {https://arxiv.org/abs/2602.23881}
}