LFM2.5-DSpark Hits 3.18x GPU Speedup With Zero Output Change

Liquid AI released DSpark draft model checkpoints for three members of its LFM2.5 family on August 20, 2026: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. The addition grafts a speculative decoding path onto each target model, delivering up to 3.18× throughput improvement on a single H100 80 GB GPU and up to 2.87× on an M4 Max MacBook Pro — with zero change to greedy output quality. For teams already tracking best small language models on Hugging Face, this release adds a concrete inference-cost lever that works from datacenter to laptop without retraining the target.

LLM decoding is memory-bound: latency is dominated by streaming weights from DRAM into SRAM, not by arithmetic. Speculative decoding attacks that bottleneck by having a cheap draft model propose candidate tokens and letting the target verify a whole block in one forward pass, amortizing weight-loading cost across every token in the block.

DSpark Architecture

The DSpark method (arXiv 2607.05147, published July 6) combines three mechanisms: a DFlash-style parallel backbone conditioned on the target model's context features; a lightweight sequential head modeled as a Markov chain between neighboring tokens to raise acceptance rates at later positions; and a confidence-scheduled verifier that predicts each token's survival probability and prunes low-confidence suffixes when verification cost would exceed the savings.

Liquid AI's implementation uses 5-layer attention-only draft models with a block size of 9, trained for 15 epochs on a mix of SFT, chat, code, and function-calling data, selecting the checkpoint with the highest acceptance rate rather than the lowest loss. The resulting drafters are approximately 300M parameters each:

Benchmark Results

All benchmarks use a DSpark block size of 9, batch size of 1, and temperature of 0. GPU measurements run on a single H100 80 GB in BF16 using SGLang; on-device measurements use llama.cpp with experimental Metal kernels on an M4 Max MacBook Pro in FP16 GGUF, with up to 256 output tokens.

The 2.6B model averages a 2.67× H100 speedup (323 → 864 tok/s) and a 2.27× MacBook speedup (61 → 139 tok/s). MATH500 yields the highest H100 gain at 3.06× (326 → 1,000 tok/s). For function-calling workloads, DSpark cuts latency by 57% on average for LFM2.5-2.6B across multi-tool scenarios — directly relevant to agentic pipelines of the kind described in what the agentic era means for data science.

The 8B MoE model (LFM2.5-8B-A1B) achieves the headline 3.18× GPU speedup on MATH500 (428 → 1,362 tok/s) but only an 18% average improvement on-device. Liquid AI attributes this gap to the current MoE implementation in llama.cpp's Metal backend: verifying a block of k tokens activates more experts and therefore generates more weight traffic than a single baseline decode step.

The 1.2B model shows the highest variance: acceptance rates range from 3.90 to 6.02 out of 10 across the five datasets, producing speedups that differ by as much as 52% depending on text distribution. GSM8K and MT-Bench return only 1.67× and 1.66× on the H100 respectively, while MATH500 reaches 2.56×. On the MacBook, the same model averages 2.54× (138 → 350 tok/s), with HumanEval reaching 2.87×.

Integration

Both framework integrations are upstreamed. SGLang support requires the build from PR #31041; llama.cpp support requires PR #27383. The SGLang server is launched by appending --speculative-algorithm DSPARK, --speculative-draft-model-path, and --speculative-draft-attention-backend flashinfer to a standard sglang.launch_server invocation, then querying the OpenAI-compatible endpoint at the configured port. The block size is read from the draft's config.json. For llama.cpp, the --spec-type draft-dspark flag enables the path, with --spec-draft-n-max 10 as the ceiling (clamped to sidecar metadata). Checkpoints are available on Hugging Face in both Safetensors and GGUF format.

Verification is exact: every proposed token is checked by the target, so greedy output is bit-identical to running the target alone. The release establishes a template for attaching lightweight speculative drafters to an existing model family without touching target weights. The 57% function-calling latency reduction for the 2.6B model, combined with 139 tok/s average on consumer Apple Silicon, meaningfully shifts what edge-deployed agents can sustain interactively. Closing the MoE on-device gap depends on Metal backend improvements to llama.cpp's expert-routing implementation.