Liquid AI released DSpark draft model checkpoints for its LFM2.5 family (1.2B-Instruct, 2.6B, and 8B-A1B), adding a speculative decoding path that speeds up inference without changing output quality. DSpark combines a DFlash-style parallel backbone, a lightweight Markov-chain sequential head, and a confidence-scheduled verifier. Reported gains include up to 3.18x throughput on H100 GPUs and up to 2.87x on an M4 Max MacBook via llama.cpp Metal, plus a 57% average cut in function-calling latency for LFM2.5-2.6B. Since verification guarantees output identical to baseline greedy decoding, benchmark accuracy is unchanged. Day-one support ships for llama.cpp and SGLang, with checkpoints available in Safetensors and GGUF formats, and integration code open-sourced upstream.
Table of contents
How does DSpark workTraining and ArchitectureQuality parityInference Speed Up on CPU and GPUHow to use LFM2.5-DSparkGet StartedCitationQuestions this post answers
What speedup does DSpark speculative decoding give for LFM2.5-2.6B inference?
DSpark draft model checkpoints deliver up to 3.18x throughput improvement on an H100 GPU and up to 2.87x on-device speedup, with LFM2.5-2.6B averaging a 2.67x speedup on H100 and 2.27x on an M4 Max MacBook Pro across five benchmarks (MATH500, HumanEval, MBPP, GSM8K, MT-Bench). It also cuts function-calling latency by 57% on average, while output remains identical to baseline greedy decoding. Engineers optimizing LLM serving costs can track speculative decoding advances like this via daily.dev.
Does speculative decoding with DSpark change LLM output quality or accuracy?
No, output quality is unchanged because DSpark speculative decoding under greedy decoding only accepts a draft token if it matches the target model's distribution; rejected tokens are replaced by the target model's own token. The emitted sequence is therefore identical to baseline greedy decoding by construction, so benchmark accuracy such as pass@1 or exact match stays the same. Anyone evaluating whether a speedup trades away accuracy can verify these trade-offs on daily.dev.
How do I run LFM2.5-2.6B with DSpark speculative decoding in SGLang?
Launch SGLang with the target model path set to LiquidAI/LFM2.5-2.6B, set --speculative-algorithm to DSPARK, point --speculative-draft-model-path to LiquidAI/LFM2.5-2.6B-DSpark, use flashinfer as the draft attention backend, and disable the radix cache with --disable-radix-cache --mem-fraction-static 0.75. This requires an SGLang build with DSpark support for LFM2 targets from PR #31041. Developers wiring up speculative decoding configs can find setup details like this through daily.dev.