# Liquid AI — LFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacBook

- Company: Liquid AI (liquid.ai)
- Announced: 2026-08-20
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.liquid.ai/blog/lfm2.5-dspark
- Record: https://forck.live/items/16771-lfm2-5-dspark-up-to-3-2x-faster-inference-from-h100-to-macbook
- Subject: LFM / d1 models
- Models affected: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, LFM2.5-8B-A1B

Liquid AI released DSpark draft model checkpoints for three LFM2.5 models, enabling speculative decoding that achieves up to 3.18x throughput improvement on GPU and up to 2.87x on-device without changing output quality. The draft models are available on Hugging Face, with integrations open-sourced in llama.cpp and SGLang.

## Evidence

Verbatim from https://www.liquid.ai/blog/lfm2.5-dspark:

> The draft models reach up to 3.18 throughput improvement on a GPU and up to 2.87x on-device.

---

Record: https://forck.live/items/16771-lfm2-5-dspark-up-to-3-2x-faster-inference-from-h100-to-macbook
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md