Try LFM • Docs • Discord

LFM2.5-24B-A2B-DSpark

LFM2.5-24B-A2B-DSpark is a speculative-decoding draft model for LiquidAI/LFM2.5-24B-A2B. It accelerates generation by proposing tokens that the target model verifies.

On NVIDIA B300 with the FP8 target in SGLang, output throughput averages 2.4× the target-only baseline at concurrency 1 (up to 3.1×), and 1.7× at concurrency 256.

Find more information about LFM2.5-DSpark in our blog post.

🗒️ Model Details

• Target model: LiquidAI/LFM2.5-24B-A2B
• Draft parameters: 327.7M (BF16)
• Backbone: 5 full attention layers, hidden_size=2048, intermediate_size=6144 with SiLU/SwiGLU, GQA with num_attention_heads=32 / num_key_value_heads=8, head_dim=64
• Extra heads: Markov head (rank 256) + confidence head
• Block size: 9
• Vocabulary: 128,000

Other models in the LFM2.5-DSpark family:

Drafter	Target
LFM2.5-1.2B-Instruct-DSpark	LFM2.5-1.2B-Instruct
LFM2.5-2.6B-DSpark	LFM2.5-2.6B
LFM2.5-VL-3B-DSpark	LFM2.5-VL-3B
LFM2.5-8B-A1B-DSpark	LFM2.5-8B-A1B
LFM2.5-24B-A2B-DSpark	LFM2.5-24B-A2B

📊 Performance

The target verifies every proposed token. Floating-point differences between batched verification and plain decoding can produce different text, including under greedy sampling. These measurements assess performance, not task accuracy. See LiquidAI/LFM2.5-24B-A2B for quality benchmarks.

Output throughput with DSpark relative to the target alone, on the full evaluation split of each dataset:

Dataset	Tokens / step	Concurrency 1	Concurrency 32	Concurrency 256
MATH-500	5.61	2.63×	3.04×	1.79×
GSM8K	4.46	1.71×	2.56×	2.15×
HumanEval	4.61	2.23×	2.63×	1.64×
MBPP	6.70	3.08×	3.28×	1.64×
MT-Bench	5.77	2.66×	2.47×	1.81×
SimpleQA	5.22	2.48×	2.91×	1.95×
LiveCodeBench	4.02	1.87×	2.20×	1.34×
AIME	4.90	2.35×	2.45×	1.36×
Arithmetic mean	5.16	2.38×	2.69×	1.71×

Each measured shard uses one NVIDIA B300, the FP8 target, a BF16 KV cache, and temperature 0. Throughput is reported per GPU; sharded results pool output tokens and measured shard seconds. Tokens per step are measured at concurrency 1 and include the target's bonus token (the ceiling is 10). The raw results and full protocol are in benchmarks/.

🏃 How to run (SGLang)

Requires a recent SGLang build with LFM2-MoE DSpark support (PR #31041) and native Hugging Face configuration support (PR #43939). Launch the target with the draft attached:

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

Then query the OpenAI-compatible endpoint at http://localhost:30000/v1. The baseline is the same command without the --speculative-* flags.

📬 Contact

• Got questions or want to connect? Join our Discord community
• If you are interested in custom solutions with edge deployment, please contact our sales team.

Citation

@article{liquidAI202624BA2B,
  author  = {Liquid AI},
  title   = {LFM2.5-24B-A2B: Built for Fast On-Device Agents},
  journal = {Liquid AI Blog},
  year    = {2026},
  note    = {www.liquid.ai/blog/lfm2-5-24b-a2b},
}

@article{liquidAI2026dspark,
  author = {Liquid AI},
  title = {LFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacBook},
  journal = {Liquid AI Blog},
  year = {2026},
  note = {www.liquid.ai/blog/lfm2.5-dspark},
}