# fal and Fireworks release EAGLE-4 analysis and prompt‑enhancement benchmark

> The joint paper introduces EAGLE‑4 speculative decoding and a 400‑item dataset for testing prompt expansion in image and video generation.

Oossa · 2026-10-21 · https://oossa.com/en/fal-and-fireworks-release-eagle-4-analysis-and-prompt-enhancement-benchmark

fal AI and Fireworks have published a technical deep‑dive of their new EAGLE‑4 speculative decoding system. The blog also releases Prompt Enhancement SpecBench, a public benchmark with 400 prompts across text‑to‑image, image‑to‑image, text‑to‑video and image‑to‑video tasks. In their tests on NVIDIA B300 GPUs, the EAGLE‑4 Harness achieved 660.6 tokens per second at batch size 1 on the Fireworks benchmark, a 1.63× speed‑up over the next best method. The tree‑based verification also raised acceptance length by up to 26 % on creative prompts. The work aims to cut latency for prompt‑enhancement steps that now dominate the end‑to‑end time of fast image and video models.

## The facts

- EAGLE‑4 Harness reaches 660.6 tokens/sec on Fireworks Bench at batch size 1 (1.63× faster than the best alternative).
- Prompt Enhancement SpecBench contains 400 examples from 18 datasets, split into four categories of 100 each.

## Why it matters

Faster prompt expansion lets developers keep interactive response times as image and video generators become near‑real‑time.

## Sources & references

1. [EAGLE-4 Technical Deepdive](https://blog.fal.ai/eagle-4-technical-deepdive/) – fal AI, 2026-10-20

Last updated: 2026-10-21