Liquid AI’s LFM2.5-24B-A2B-DSpark card reports eight workloads

Liquid AI’s public Hugging Face model card identifies LFM2.5-24B-A2B-DSpark as a draft model for LFM2.5-24B-A2B and reports throughput results across eight datasets. The figures compare output throughput with the target model running alone. OrcaRouter reported that the repository appeared on October 19, 2026.

How the draft model works with its target

LFM2.5-24B-A2B-DSpark is designed for speculative decoding, a method in which a smaller draft model proposes tokens and a larger target model verifies them. The drafter supports LFM2.5-24B-A2B; it does not replace it. The model card lists 327.7 million BF16 parameters and a block size of 9.

Liquid AI’s reported throughput by workload

Liquid AI’s results vary by dataset and concurrency. The table gives the reported throughput ratio against the target model alone; for example, MBPP reaches 3.28× at concurrency 32, while LiveCodeBench is at 1.34× at concurrency 256.

The benchmark uses one NVIDIA B300, an FP8 target, a BF16 KV cache and temperature 0. Throughput is measured per GPU across each dataset’s full evaluation split. “Tokens per step” is measured at concurrency 1, includes the target’s bonus token and has a stated ceiling of 10. The spread across workloads means the arithmetic mean does not describe every dataset equally: at concurrency 256, for instance, GSM8K’s reported ratio is 2.15× while LiveCodeBench’s is 1.34×.

What the performance figures measure

Liquid AI says the measurements assess performance, not task accuracy. It also warns that floating-point differences between batched verification and plain decoding can produce different text, including with greedy sampling. The listed results are from the B300 setup, not measurements of the 24B drafter on Apple silicon.

SGLang setup described in the model card

Liquid AI’s SGLang instructions pair the FP8 target with the DSpark drafter, set the draft block size to 9 and request 10 draft tokens. They require a recent SGLang build with LFM2-MoE DSpark support and native Hugging Face configuration support. The model card gives this launch command:

python -m sglang.launch_server \
  --model-path LiquidAI/LFM2.5-24B-A2B-FP8 \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path LiquidAI/LFM2.5-24B-A2B-DSpark \
  --speculative-draft-attention-backend flashinfer \
  --speculative-dspark-block-size 9 \
  --speculative-num-draft-tokens 10 \
  --mem-fraction-static 0.8 --port 30000