HomeAILiquid AI Releases DSpark Draft Models
AI

Liquid AI Releases DSpark Draft Models for LFM2.5

Checkpoints for three LFM2.5 models add speculative decoding support in llama.cpp and SGLang, speeding up inference by up to 3.18x.

WHAT YOU NEED TO KNOW
  • Liquid AI published ~300M-parameter DSpark draft models for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B on Hugging Face.
  • Throughput improved by up to 3.18x on an Nvidia H100 GPU and up to 2.87x on an Apple M4 Max chip.
  • Checkpoints include day-one upstream integration support for SGLang and llama.cpp in Safetensors and GGUF formats.

Liquid AI released DSpark draft model checkpoints on Hugging Face for three language models in its LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. The release introduces speculative decoding to accelerate token generation while preserving baseline greedy output accuracy.

Speculative decoding tackles memory bandwidth constraints during text generation by pairing a lightweight draft model with a larger target model. The draft network proposes candidate tokens that the target model validates in a single forward pass. DSpark incorporates a parallel backbone derived from DFlash, a sequential Markov-chain head linking adjacent tokens, and a confidence-scheduled verifier that prunes low-confidence token suffixes.

Each draft checkpoint contains around 300 million parameters across a five-layer, attention-only architecture with a block size of nine. Liquid AI trained the draft models over 15 epochs on a dataset containing supervised fine-tuning, chat, programming code, and tool-use examples. The team selected checkpoints based on peak acceptance rates rather than minimal training loss.

Benchmark results

Hardware evaluations conducted on an Nvidia H100 80 GB graphics card and an Apple M4 Max MacBook Pro demonstrated throughput gains across five evaluation datasets. Tests on the H100 using SGLang showed speedups reaching up to 3.18x for LFM2.5-8B-A1B, lifting throughput from 428 to 1,362 tokens per second. On the M4 Max laptop running llama.cpp and Metal kernels, LFM2.5-1.2B-Instruct achieved up to a 2.87x speedup, moving from 136 to 389 tokens per second.

LFM2.5-2.6B reached an average latency reduction of 57 percent in multi-tool function-calling tests, delivering roughly 140 tokens per second on the M4 Max system. In contrast, the mixture-of-experts model LFM2.5-8B-A1B recorded an average on-device throughput gain of 18 percent on Apple silicon. Liquid AI attributed this lower edge speedup to llama.cpp's current Metal MoE implementation and increased memory traffic from activating multiple experts during multi-token verification.

Deployment options

Hugging Face hosts the draft models in both Safetensors and GGUF formats. Users can run the checkpoints via SGLang server commands or through llama-server with speculative draft flags.

Xentir Media
Xentir Media NewsroomSource-backed AI and technology coverage, drafted by Xentir's automated editorial system under fixed human-set rules. See our editorial policy and AI usage policy.
J
Jomon · Founder & EditorFounder and editor of Xentir Media. Sets the editorial rules the newsroom system runs under, and is accountable for its corrections. About Jomon · [email protected]
The Xentir Brief
The developments worth knowing — one useful email.
Get the Brief →