Liquid AI Releases DSpark Draft Model, Speeding Decoding up to 3.18x

On August 21, Liquid AI released DSpark, a 300M-parameter draft model for the LFM2.5 family, introducing a speculative decoding path with measured throughput gains.

Confirmed

• DSpark is a draft model (300M parameters) for the LFM2.5 series using speculative decoding
• Mechanism: a lightweight draft model proposes candidate token blocks, validated by the target model in a single forward pass, trading minimal memory for significant decoding speedup with unchanged output quality
• Official benchmarks: up to 3.18x decoding speedup on H100 (MATH500); M4 Max was also tested though specific figures were not given in the posts

Why it matters

• Speculative decoding is a mainstream approach to cutting inference costs; DSpark demonstrates a viable path pairing a small draft model with the LFM2.5 main model
• Roughly tripling throughput with near-zero quality loss and minor extra memory has direct implications for real-world deployment costs of LFM2.5

2026-08-21 ~ 2026-08-21 · 5 related posts

• [source] Liquid AI Releases DSpark: Speculative Decoding Up to 3.18x Faster — helloiamleonie · 2026-08-21
• Liquid AI releases DSpark draft models, speeding up decoding by up to 4x — JosephJacks_ · 2026-08-21
• Liquid AI Releases DSpark Draft Models for Up to 3.18x Faster Decoding — JosephJacks_ · 2026-08-21

2 near-duplicate retellings: JosephJacks_ · JosephJacks_