Liquid AI Releases DSpark Draft Models for Faster LFM2.5 Inference

In a blog post published August 20, 2026 on Hugging Face, Liquid AI announced the release of DSpark draft model checkpoints for three models in its LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. According to the post, the checkpoints add a speculative decoding path that increases decoding speed with a minimal memory increase and no change to output quality. Speculative decoding uses a lightweight draft model to produce candidate tokens, which the target model verifies in a single forward pass. Liquid AI states its draft models are attention-only, with 5 layers and around 300 million parameters each, trained for 15 epochs. The company reports that because a draft token is only accepted if it matches the target model's distribution under greedy decoding, benchmark accuracy is unchanged. Liquid AI measured throughput using llama.cpp with Metal on an M4 Max MacBook Pro and SGLang on a single H100 80GB GPU. It reports that DSpark reduces latency by 57% on average for LFM2.5-2.6B in multi-tool scenarios, up to 52% variance for LFM2.5-1.2B-Instruct, and an 18% average improvement on-device for LFM2.5-8B-A1B, which it attributes to the current MoE implementation in llama.cpp's Metal backend. The checkpoints are available on Hugging Face in Safetensors and GGUF formats.

• Draft models released for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B on August 20, 2026
• Each draft model is attention-only with 5 layers and roughly 300M parameters
• Reported average latency reduction of 57% for LFM2.5-2.6B in multi-tool scenarios; up to 3.2x faster overall
• Output quality unchanged because the target model verifies every proposed token under greedy decoding
• Ships with day-one support for llama.cpp (Metal) and SGLang; available in Safetensors and GGUF