Home/Events/Up to 3.2x Faster Inference with LFM2.5-DSpark

Up to 3.2x Faster Inference with LFM2.5-DSpark

Confirmed

Confidence

90%

Impact: 80%

Updated 4h ago

Consensus Brief

Hugging Face has released draft model checkpoints for three models in the LFM2.5 family, which include LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. These models utilize a new speculative decoding method called DSpark, resulting in significant improvements in inference speed without compromising output quality.

Sourced from

Hugging FaceT1

Primary: Hugging Face

What Changed Since Last Update

4h ago

The introduction of DSpark allows for up to 3.2x faster inference on GPUs and 2.87x on-device, along with a 57% reduction in function-calling latency for the LFM2.5-2.6B model.

Claim Ledger

4 claims tracked across sources

Confirmed Fact

DSpark achieves up to 3.2x throughput improvement on a GPU and up to 2.87x on-device.

Confirmed Fact

Cuts function-calling latency by 57% on average for LFM2.5-2.6B.

Confirmed Fact

The draft models are relatively small, with each around ~300M parameters.

Confirmed Fact

Day-one support for llama.cpp and SGLang is provided.

Role-Based Impact Analysis

Source Timeline

1 source corroborating

Up to 3.2x Faster Inference with LFM2.5-DSpark

Hugging Face·4h ago