LiquidAI released DSpark draft model checkpoints for three models in its LFM2.5 family on August 20, marking the first public release of speculative decoding models for Liquid Foundation Models1. The release covers LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B.
The DSpark draft models add a speculative decoding path that trades a minimal memory increase for a large decoding speedup without changing output quality. On a GPU, the draft models reach up to 3.18x throughput improvement; on-device, they reach up to 2.87x. Under greedy decoding, benchmark accuracy is unchanged because the emitted sequence is identical to baseline greedy by construction.
Each draft model is approximately 300M parameters. LiquidAI measured GPU throughput with SGLang on a single Nvidia H100 80 GB in BF16 at batch size 1 and temperature 0. On-device throughput was measured with llama.cpp and Metal on an M4 Max MacBook Pro using FP16 GGUF weights at batch size 1, temperature 0, and up to 256 output tokens.
Benchmark-specific results show LFM2.5-2.6B achieved a 3.06x speedup on Nvidia H100 on MATH500 and a 2.87x speedup on MT-Bench. Across various multi-tool scenarios, DSpark reduces BFCL function call latency by 57% on average. Results vary by model and workload: for LFM2.5-1.2B-Instruct, a non-reasoning model, speedup varies by as much as 52% depending on the distribution of the underlying text. For LFM2.5-8B-A1B, the observed real-world speedup on edge devices is only an 18% improvement.
LiquidAI trained each draft model and ran all ablation studies exclusively on AMD hardware using its own training framework. The final published draft checkpoints were selected using the epoch with the highest acceptance rate rather than based on loss.
The DSpark draft models are available on Hugging Face. The LFM-compatible DSpark integration is open-sourced upstream in llama.cpp and SGLang, with day-one support across both frameworks. Throughput results were evaluated on five benchmark datasets: MATH500, GSM8K, HumanEval, MBPP, and MT-Bench.
LiquidAI stated it believes co-designing model architecture with speculation methods to model real-world inference characteristics will be a crucial part of model design going forward.
ANALYSIS The wide variance in speedup across models and hardware — from 18% on-device for LFM2.5-8B-A1B to over 3x on GPU for LFM2.5-2.6B — underscores that speculative decoding gains are workload- and architecture-dependent, not uniform. The exclusive use of AMD hardware for training, rather than NVIDIA CUDA, is a notable infrastructure choice for a company releasing models benchmarked primarily on Nvidia H100.