Liquid AI's DSpark drafters hit 2x decode speed on iPhone within 24 hours
helloiamleonie · x · 2026-08-21
Less than 24 hours after Liquid AI released DSpark drafters (draft models for speculative decoding), users are running LFM2.5 on iPhones with roughly 2x faster decoding. A Japanese developer reported that following the Liquid AI team's advice to load in F16 precision markedly sped up decode; Liquid AI staff amplified the community result.
More from Infra
- New Sparse Attention Nodes for H3 Model Boost Speed by 30% — Zironic · 2026-08-21
- Hugging Face Adds NVFP4 Kernel Support, Cutting VRAM by 61% — ariG23498 · 2026-08-21
- Dev Observes: OSS Inference Providers Lag Behind OpenAI in Reliability — NielsRogge · 2026-08-21
- NVIDIA's Aggressive NeoCloud Moves Ahead of Earnings Report — RihardJarc · 2026-08-21
- Neoclouds Are AI’s New Power Brokers, But Enterprises Should Be Careful — DavidLinthicum · 2026-08-21
- DevOps Interview: Using CloudTrail to Audit AWS Changes — _jaydeepkarale · 2026-08-21