#speculative decoding

2 articles tagged

Chart-style graphic comparing LFM2.5-DSpark speedups on H100 and Apple M4 Max, August 2026

Engineering · Aug 22, 2026 · 12 min read

LFM2.5-DSpark ships 3.2x on an H100 and 1.18x on a MacBook: the 20 August 2026 numbers mobile teams should read

Liquid AI released LFM2.5-DSpark speculative-decoding drafters on 20 August 2026. The headline 3.2x is a single H100 cell. On an M4 Max the 2.6B mean is 2.27x and the 8B MoE mean is 1.18x.

Manu Shukla

#on-device AI #speculative decoding #LFM2.5

vLLM parallel speculative decoding: P-EAGLE, DFlash, DSpark hero

AI Tools · Aug 3, 2026 · 13 min read

vLLM Speculative Decoding in 2026: P-EAGLE vs DFlash vs DSpark for Faster, Cheaper LLM Serving

P-EAGLE, DFlash, and DSpark bring parallel drafting to vLLM in 2026, generating a whole block of draft tokens in one pass for up to 1.69x higher throughput than EAGLE-3. How they work and how to serve them.

Manu Shukla

#vLLM #speculative decoding #LLM inference