Liquid AI Releases DSpark to Speed Up LFM2.5 Models

Liquid AI has introduced DSpark draft model checkpoints for three of its LFM2.5 models: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. This release introduces speculative decoding, where a smaller draft model of roughly 300 million parameters proposes a block of nine candidate tokens that the larger target model verifies in a single forward pass. The draft model for LFM2.5-1.2B-Instruct uses 295.7 million parameters, while the drafters for LFM2.5-2.6B and LFM2.5-8B-A1B use 327.7 million parameters. The architecture features five full-attention layers, a hidden size of 2048, an intermediate size of 6144, and grouped-query attention with 32 heads over 8 key-value heads.

The upgrade yields significant speedups with zero change to the final output under greedy decoding, preserving benchmark accuracy. On an Nvidia H100 GPU running BF16 via SGLang, the LFM2.5-1.2B-Instruct model saw its mean throughput jump from 656 to 1384 tokens per second (a 2.10x speedup), peaking at 2.56x on MATH500. The LFM2.5-2.6B model rose from 323 to 864 tokens per second (2.67x mean speedup), and the LFM2.5-8B-A1B model went from 418 to 1074 tokens per second (2.54x mean speedup), reaching a 3.18x peak on MATH500. On an Apple M4 Max MacBook Pro running llama.cpp, the 1.2B model achieved a 2.54x mean speedup (138 to 350 tokens per second), while the 2.6B model reached 2.27x (61 to 139 tokens per second). However, the 8B-A1B model on the M4 Max only saw a 1.18x mean speedup due to current Metal backend limitations for mixture-of-experts models.

For developers, this release dramatically reduces latency in local applications. In multi-tool function-calling scenarios, DSpark cuts latency by an average of 57% for LFM2.5-2.6B. The weights are available as Safetensors and GGUF for self-hosting, with day-one support in SGLang and llama.cpp. The 2.6B drafter repository adds a memory cost of 655 megabytes in BF16. Under the LFM Open License v1.0, commercial use is free for entities with under $10 million in annual revenue, while larger enterprises must obtain a commercial license.