Try LFM • Docs • LEAP • Discord

LFM2.5-1.2B-Instruct-DSpark

LFM2.5-DSpark is a family of speculative-decoding draft models that adapt DSpark for the LFM2.5 architecture. They allow LFM2.5 models to run faster without degrading quality.

This is a drafter for LiquidAI/LFM2.5-1.2B-Instruct. In SGLang, decoding runs about 2× faster. It also runs on-device on Apple silicon through the Metal backend.

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

🗒️ Model Details

LFM2.5-1.2B-Instruct-DSpark is a DSpark speculative-decoding draft model with the following features:

Target model: LiquidAI/LFM2.5-1.2B-Instruct
Draft parameters: 295.7M (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: 65,536

Other models in the LFM2.5-DSpark family:

Drafter	Target
LFM2.5-1.2B-Instruct-DSpark	LFM2.5-1.2B-Instruct
LFM2.5-8B-A1B-DSpark	LFM2.5-8B-A1B
LFM2.5-2.6B-DSpark	LFM2.5-2.6B

📊 Performance

Benchmarks

Speculative decoding is exact: the target verifies every proposed token, so the generated text is what the target would have produced on its own. See LiquidAI/LFM2.5-1.2B-Instruct for performance benchmarks.

Acceptance

Mean accepted tokens per decoding step, by benchmark (1×H100, batch size 1, greedy decoding). Higher means more of the draft's proposed block is accepted per target forward pass, so decoding is faster (at block size 9, the ceiling is 10).

Benchmark	Accepted tokens / step
MATH-500	5.78
GSM8K	4.25
HumanEval	5.51
MBPP	5.41
MT-Bench	3.11
Mean	4.81

On-device and GPU Inference

Dataset	Acceptance (of 10)	Speedup on H100	Speedup on M4 Max
MATH500	6.02	2.56x 668 → 1712 tok/s	2.62x 140 → 366 tok/s
HumanEval	5.31	2.26x 664 → 1499 tok/s	2.87x 136 → 389 tok/s
MBPP	5.52	2.37x 667 → 1578 tok/s	2.74x 137 → 375 tok/s
GSM8K	4.34	1.67x 624 → 1041 tok/s	2.73x 140 → 381 tok/s
MT-Bench	3.90	1.66x 657 → 1091 tok/s	1.72x 137 → 237 tok/s
Mean	5.02	2.10x 656 → 1384 tok/s	2.54x 138 → 350 tok/s

🏃 How to run (SGLang)

Requires a build of SGLang with DSpark support for LFM2 targets (PR #31041). Launch the target with the drafter attached:

python -m sglang.launch_server \
  --model-path LiquidAI/LFM2.5-1.2B-Instruct \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path LiquidAI/LFM2.5-1.2B-Instruct-DSpark \
  --speculative-draft-attention-backend flashinfer \
  --disable-radix-cache --mem-fraction-static 0.75 --port 30000

Then query the OpenAI-compatible endpoint at http://localhost:30000/v1. The block size is read from the draft's config.json; the baseline is the same command without the three --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{liquidAI202626B,
  author  = {Liquid AI},
  title   = {LFM2.5-2.6B: Agents Everywhere},
  journal = {Liquid AI Blog},
  year    = {2026},
  note    = {www.liquid.ai/blog/lfm2-5-2-6b},
}

@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},
}