Liquid AI releases LFM2.5-DSpark, open-weight draft models for up to 3.2x faster inference

Liquid AI has released LFM2.5-DSpark, a family of open-weight speculative-decoding draft models built to speed up inference on its LFM2.5 model line without changing output quality.

What's new

Per the model card on Hugging Face: "LFM2.5-DSpark is a family of speculative-decoding draft models that adapt DSpark for the LFM2.5 architecture." Liquid AI shipped three draft models pairing with LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B, each a small companion model (327.7M parameters for the 8B-A1B drafter, in BF16 precision, with 5 full attention layers) that proposes candidate tokens for the larger target model to verify in parallel rather than generating one token at a time.

The measured speedups vary by hardware and model. On H100 GPUs running SGLang, the 8B-A1B pairing sees a mean 2.54x speedup (418 to 1,074 tokens/second), with a peak of 3.18x on the MATH-500 benchmark. On-device, on an Apple M4 Max, the same pairing sees a more modest 1.18x mean speedup (90 to 106 tokens/second). The draft model's mean acceptance rate — how many proposed tokens the target model actually accepts per decoding step — is 7.21 across benchmarks, ranging from 3.91 on GSM8K to 8.99 on MT-Bench. The models ship with support for both llama.cpp, for edge and on-device inference, and SGLang, for GPU-accelerated serving.

Context

Speculative decoding is an increasingly standard technique for cutting inference latency: a small, fast draft model proposes several tokens ahead, and the larger target model verifies them in one pass instead of computing each token sequentially, with acceptance rate determining how much of that speedup is actually realized. Liquid AI has been iterating quickly on the LFM2.5 line this month, having released LFM2.5-VL-3B, an open-weight vision-language model, on August 12; DSpark is a serving-side companion release rather than a new base model, aimed specifically at making the existing LFM2.5 family cheaper and faster to run in production.

Why it matters

The hardware-dependent results are the most informative part of this release: a 2.5-3.2x speedup on datacenter GPUs but under 1.2x on Apple silicon shows that speculative decoding gains aren't free or uniform — they depend heavily on how the target architecture (here, a mixture-of-experts model) maps onto a given inference stack, with Liquid AI itself attributing the smaller on-device gain to how MoE models currently run on Metal in llama.cpp. For teams evaluating LFM2.5 for latency-sensitive deployment, that's a concrete, hardware-specific number to plan around rather than a generic inference-speedup marketing claim.