Sakhanda Wire
NVDA $150.96 -2.89% MSFT $508.29 +0.40% GOOGL $282.44 +0.80% META $616.38 -0.40% AMZN $267.74 +3.82%
← Back to the news

Liquid AI releases DSpark models for faster LFM2.5 decoding

Liquid AI has released DSpark draft checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B. Each checkpoint adds a roughly 300M-parameter drafter that proposes nine tokens, while the target model verifies them in one forward pass. Under greedy decoding, the generated sequence remains identical to the target model’s output alone, so benchmark accuracy is unchanged.

Liquid AI reports speedups of up to 3.18x on an H100 and up to 2.87x on an M4 Max MacBook Pro. Results depend on token acceptance and workload: LFM2.5-8B-A1B reached 3.18x on MATH500 but averaged 1.18x on the M4 Max. The company attributes the weaker Apple-silicon result to the current Metal implementation for mixture-of-experts models and increased weight traffic during verification.

The checkpoints are available in Safetensors and GGUF formats, with support from llama.cpp and SGLang. They require a compatible self-hosted build and are not currently served by a hosted Hugging Face inference provider. Liquid AI also reports 57% lower latency for multi-tool function calling with LFM2.5-2.6B. Commercial use is free under the LFM Open License v1.0 for entities below $10M in annual revenue; larger companies must obtain a commercial license.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news