Skip to content
Other

qwen2.5-3b-instruct-eagle3

speeding up Qwen2.5-3B generation in vLLM

2 min readupdated

A lightweight draft head for speculative decoding (a speedup method where a small model predicts upcoming tokens and the main model quickly verifies them) targeting Qwen/Qwen2.5-3B-Instruct. Built by ressf on the EAGLE-3 architecture, this 0.47B-parameter adapter uses a single decoder layer and a 32,000-token vocabulary. Unlike running the base model standalone, this add-on boosts token generation speed by up to 22% under light load while keeping output quality virtually untouched. It is a solid fit for developers deploying Qwen2.5-3B-Instruct on local or server GPUs with low concurrent traffic.

Key features

  • +18% to +22% generation speedup across 1 to 16 concurrent clients (benchmarked on an RTX 5070 Ti 16 GB in FP16).
  • Trained via the speculators library on 9,618 UltraChat dialogues regenerated by Qwen2.5-3B-Instruct itself.
  • Mean acceptance length of 1.52 tokens, peaking when predicting 1 speculative token per step.
  • Subsample accuracy stays close to baseline: 0.580 vs. 0.585 on both GSM8K and IFEval.

What you can do

  • Speed up a local vLLM 0.29.0 inference server for chatbots or assistants handling low concurrency.
  • Use it as a ready-made reference and baseline for speculative decoding experiments paired with TRITON_ATTN.

Example

VLLM_USE_V2_MODEL_RUNNER=0 vllm serve Qwen/Qwen2.5-3B-Instruct --dtype half --attention-backend TRITON_ATTN \
  --speculative-config '{"method":"eagle3","model":"","num_speculative_tokens":1,"attention_backend":"TRITON_ATTN"}'
Read moreDetails
Speed (output tok/s on RTX 5070 Ti 16 GB): +20% (1 client), +22% (2–8 clients), +18% (16), +11% (32), −4% (64).
Benchmarks (with head / without): GSM8K 0.580 / 0.585, IFEval 0.580 / 0.585, MMLU-Pro 134.4 / 136.7 out of 350.
Limitations: runs slower than the base model at high batch sizes (64 clients) and on the FP8-quantized model at 32–64 clients.
License: Qwen Research License (includes the base model's embedding matrix).
Discuss Telegram 0 22

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.9 GB · ready files1.5 GB
  • FP8 / INT8recommendedweights 0.4 GB1.0 GB
  • 4-bitweights 0.2 GB0.8 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # ressf/qwen2.5-3b-instruct-eagle3
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download ressf/qwen2.5-3b-instruct-eagle3 --local-dir ~/models/qwen2.5-3b-instruct-eagle3
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. Qwen2.5-3Bfamily root
  2. Qwen2.5-3B-Instruct
  3. qwen2.5-3b-instruct-eagle3

New posts about versions of Qwen2.5-3B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face
Merges 3All on Hugging Face

About the model

ressf/qwen2.5-3b-instruct-eagle3
Author
ressf
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

220

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments