SpecForge

Appearance

Production-Grade Draft Models Built by Our Community

Trained with SpecForge. Served with SGLang.

Collection on Hugging Face Performance Dashboard Usage

Models ​

Showing 1–9 of 30 matching draft models · 3.0M downloads across the collection Snapshot · updated Sep 4, 2026

Performance Dashboard ​

We evaluate SpecBundle draft models on conversation (MT-Bench), general knowledge (GPQA, FinanceQA), math (GSM8K, Math500) and coding (HumanEval, LiveCodeBench) benchmarks under different speculative decoding configurations (steps, top-k and number of draft tokens). Pick a target model to compare its draft models against the baseline without speculative decoding.

Target modelQwen3-30B-A3B-Instruct-2507
Draft models1
Configurations5
Best avg. speedup1.82×
Detailed results
Draft modelConfiggsm8kacc. len / tokens/smath500acc. len / tokens/smtbenchacc. len / tokens/shumanevalacc. len / tokens/slivecodebenchacc. len / tokens/sfinanceqaacc. len / tokens/sgpqaacc. len / tokens/sAvg speedupthroughput vs. baseline
Baseline–1.00/1071.291.00/1468.951.00/1341.351.00/1366.621.00/1492.621.00/1320.131.00/1410.431.00×
Qwen3-30B-A3B-Instruct-2507-SpecForge8-3-1-42.64/1488.361.39× 3.40/3022.302.06× 2.50/2048.691.53× 3.35/2618.171.92× 3.18/2911.411.95× 2.08/1778.971.35× 2.60/2237.791.59× 1.68×
Qwen3-30B-A3B-Instruct-2507-SpecForge8-5-1-63.01/1499.621.40× 4.50/3458.772.35× 2.83/2086.121.56× 4.38/2912.142.13× 4.02/3265.252.19× 2.27/1778.681.35× 3.01/2341.301.66× 1.81×
Qwen3-30B-A3B-Instruct-2507-SpecForge8-5-3-62.53/1491.181.39× 3.83/2710.071.84× 2.56/1698.421.27× 3.79/2367.021.73× 3.30/2455.091.64× 2.27/1652.161.25× 2.69/1961.171.39× 1.50×
Qwen3-30B-A3B-Instruct-2507-SpecForge8-7-1-83.15/1438.401.34× 5.29/3636.152.48× 2.98/1998.161.49× 5.12/3069.982.25× 4.58/3413.032.29× 2.30/1682.961.27× 3.22/2310.211.64× 1.82×
Qwen3-30B-A3B-Instruct-2507-SpecForge8-7-4-82.52/1478.341.38× 3.98/2651.001.80× 2.74/1742.981.30× 4.03/2363.341.73× 3.40/2355.091.58× 2.61/1753.671.33× 2.92/2008.741.42× 1.51×

Usage ​

Launch an SGLang server with a target model and its SpecBundle draft. Add --tp, --ep and --mem-fraction-static when you run into memory limits.

bash
python3 -m sglang.launch_server \
    --model <target-model-path> \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path <draft-model-path> \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4
bash
python3 -m sglang.launch_server \
    --model <target-model-path> \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path <draft-model-path> \
    --speculative-dflash-block-size 8

Each model card on Hugging Face documents the exact launch flags and the serving configuration used for its published benchmarks.

Released under the MIT License.

Copyright © 2025-2026 SpecForge Team · SGLang