LiquidAI released three DSpark draft model checkpoints for its LFM2.5 family—LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B—using speculative decoding to trade minimal memory increase for large inference speedups. On a single H100 GPU, the LFM2.5-8B-A1B draft model delivers up to 3.18x throughput improvement (428 → 1362 tokens/sec on MATH500), while on-device (M4 Max MacBook with Metal) speedups range from 1.18x to 2.87x depending on model size. Critically, output quality is unchanged—greedy decoding ensures emitted sequences are identical to baseline by construction.
DSpark combines three techniques: a DFlash-style parallel backbone producing hidden states for all draft tokens in a single forward pass, a lightweight sequential Markov-chain head to raise acceptance rates, and a confidence-scheduled verifier that prunes low-confidence token suffixes when verification costs exceed the savings. The draft models are ~300M parameters each with only 5 attention layers, keeping overhead minimal. For function-calling scenarios, LFM2.5-2.6B sees 57% latency reduction on average across five multi-tool datasets.
Deployment accessibility matters: day-one support for llama.cpp and SGLang means practitioners can ship these models immediately on-device and cloud without custom infrastructure. For architects targeting interactive inference and agent loops where latency dominates, these numbers hint that post-training inference optimizations (not just raw parameter scaling) are yielding material wall-clock improvements. The 2.67x mean speedup on H100 at zero output-quality cost is a data point on the inference curve.