Liquid AI released speculative-decoding draft checkpoints for three models in its LFM2.5 family on August 20, 2026, reporting throughput gains of up to 3.18x on a single H100 GPU and up to 2.87x on an Apple-silicon MacBook, with no change to model outputs. The LFM2.5-DSpark release covers drafters for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and the mixture-of-experts LFM2.5-8B-A1B, each adding roughly 300 million parameters of draft overhead on top of the target model.
The checkpoints ship in Safetensors and GGUF formats with day-one support in llama.cpp and SGLang, both integrations contributed upstream to the official codebases. Because speculative decoding only emits tokens the target model has verified, the company states the generated text is identical to what the target would produce alone under greedy decoding, so benchmark accuracy is unchanged.
Liquid AI’s measurements, run at batch size 1 and temperature 0 across five datasets, put the mean speedup for LFM2.5-2.6B at 2.67x on an H100 (323 to 864 tokens per second) and 2.27x on an M4 Max MacBook Pro (61 to 139 tokens per second). The largest single result came from LFM2.5-8B-A1B on MATH500, where throughput on the H100 rose 3.18x, from 428 to 1,362 tokens per second. The company also reports that DSpark cut function-calling latency by 57% on average for LFM2.5-2.6B across multi-tool scenarios, the headline result for the on-device agentic workloads the LFM2.5 line is aimed at.
How DSpark Speeds Up Decoding
The decode phase of LLM inference is memory-bound: most of the latency comes from streaming weights from DRAM into on-chip memory rather than from computation itself, which is why inference economics have become the field’s central engineering problem. Speculative decoding attacks this by having a small draft model propose a block of candidate tokens, then verifying the whole block in one forward pass of the target model, spreading the cost of loading the weights across every token checked.
DSpark, introduced in a July 2026 paper by DeepSeek researchers and deployed in that company’s DeepSeek-V4 serving system, combines three components: a parallel backbone that produces hidden states for all draft tokens in a single pass, a lightweight sequential head that models dependencies between neighboring tokens to keep acceptance rates from decaying late in the block, and a confidence-scheduled verifier that prunes low-confidence suffixes when verifying them would cost more than it saves. In DeepSeek’s production deployment, the paper reports per-user generation speedups of 60 to 85% over the prior MTP-1 baseline at matched throughput.
Liquid AI’s drafters follow that recipe with a simplified attention-only design: five layers, a block size of nine draft tokens per step, and a Markov head over a 128,000-token vocabulary, per the LFM2.5-2.6B-DSpark model card. Each drafter was trained for 15 epochs on a mix of supervised fine-tuning, chat, code, and function-calling data, with the checkpoint selected by highest acceptance rate rather than lowest loss. The exactness guarantee does the quality work: “Speculative decoding is exact: the target verifies every proposed token, so greedy output equals the target alone,” the GGUF model card states, with per-response timings exposing how many draft tokens were proposed and accepted.
LFM2.5-DSpark by the Numbers
- 3.18x — best reported GPU speedup (LFM2.5-8B-A1B, MATH500, H100: 428 → 1,362 tok/s)
- 2.87x — best reported on-device speedup (LFM2.5-1.2B-Instruct, HumanEval, M4 Max: 136 → 389 tok/s)
- 2.67x / 2.27x — mean H100 / M4 Max speedups for LFM2.5-2.6B across five datasets
- 57%: average function-calling latency reduction for LFM2.5-2.6B in multi-tool scenarios
- 295.7M–327.7M (draft model parameters, against targets of 1.2B to 8B)
- 4.81 of 10, mean accepted draft tokens per step for LFM2.5-2.6B at block size 9
Where the Reported Speedups Narrow
Liquid AI’s own tables show the gains are uneven, and the company states the reasons. For LFM2.5-8B-A1B, on-device improvement averages only 1.18x despite the highest acceptance rates of the three models, a gap the company attributes to the current mixture-of-exets implementation in llama.cpp’s Metal backend and to the extra weight traffic that verifying a block of tokens activates across experts. For LFM2.5-1.2B-Instruct, acceptance rates vary enough by dataset that the speedup swings by as much as 52% depending on the text distribution, from 1.66x on MT-Bench up to 2.56x on MATH500 on the H100.
All figures are vendor-reported from Liquid AI’s own harness: SGLang on one H100 80GB in BF16 for GPU numbers, llama.cpp with experimental Metal kernels on an M4 Max with FP16 GGUF weights for on-device numbers, capped at 256 output tokens. The SGLang path requires a build with DSpark support for LFM2 targets, and the llama.cpp path requires the corresponding build, so the speedups depend on those integrations rather than shipping in a stable release of either engine.
Liquid AI’s On-Device Push So Far
The DSpark release is the third LFM2.5-family update in just over a week. On August 12, 2026, the company shipped LFM2.5-VL-3B, a vision-language model for the edge, and on August 19, 2026 it published quantization-aware distilled Q4_0 checkpoints for the family. The through-line is the same: the company says the 2.6B model’s DSpark speedup on a MacBook pushes interactivity beyond the throughput offered by most proprietary cloud models, which it puts at roughly 140 tokens per second.
All three drafters are available now on Hugging Face: LFM2.5-1.2B-Instruct-DSpark, LFM2.5-2.6B-DSpark, and LFM2.5-8B-A1B-DSpark, with GGUF builds alongside for llama.cpp deployments.
Credit: Source link
