Liquid AI released DSpark draft models for the LFM2.5 family (1.2B-Instruct, 2.6B, and 8B-A1B), adding speculative decoding—a technique where a small draft model proposes token candidates that a larger target model verifies in a single pass—without changing output quality. The speedups are substantial: up to 3.18x throughput improvement on an H100 GPU and up to 2.87x on an M4 Max MacBook Pro using on-device inference. The draft models are roughly 300M parameters each, adding minimal memory overhead.
The most significant impact appears in agentic workloads where the model reasons before each tool call. Across multi-tool function-calling scenarios on the BFCL benchmark, DSpark reduces latency by 57% on average for LFM2.5-2.6B. This matters because agents incur the decode cost multiple times per turn—once for planning, once after each tool call, and again for re-planning. Speedup varies by task: MATH reasoning on H100 achieves 3.18x, while GSM8K (which requires fewer hops) yields 1.29x, tracking the model's acceptance rate.
DSpark's architecture unifies three components: a DFlash-style parallel backbone that generates all draft tokens in one forward pass (fast but initially incoherent), a lightweight Markov-chain sequential head that adds inter-token dependency to raise acceptance at later positions, and a confidence-scheduled verifier that prunes low-acceptance suffixes when verification cost exceeds the savings. The technique is exact: target verification ensures output is identical to the baseline greedy decode, so no accuracy is sacrificed.
Architects deploying local agents and on-device reasoning workflows should track this release: LFM2.5-DSpark ships with day-one support in llama.cpp and SGLang, making it immediately usable. The Liquid AI team trained drafts on diverse data (SFT, chat, code, function-calling) and selected epochs by acceptance rate rather than loss—a signal shift that hints at the real bottleneck in decode performance: not compute, but memory bandwidth and cache coherence.