Liquid AI Releases LFM2.5 Draft Models for Faster Decoding

Liquid AI has released DSpark draft model checkpoints for three models within its LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. These "drafter" models introduce a speculative decoding path to existing target models, significantly accelerating the token generation process. The core mechanism involves a smaller draft model, approximately 300 million parameters, proposing a block of nine candidate tokens. A larger, target model then verifies this entire block in a single forward pass. This approach results in a substantial decoding speedup, achieving up to 3.18x faster performance on NVIDIA H100 GPUs and up to 2.87x faster on Apple M4 Max MacBook Pro hardware. Crucially, this speed enhancement does not compromise the model's output quality. Under greedy decoding, the sequence of tokens emitted by the combined system is identical to that produced by the target model operating independently, meaning benchmark accuracy remains unchanged.

Both the llama.cpp and SGLang inference engines offer day-one support for these new draft models. Deployment is currently feasible for users who self-host their AI infrastructure. The model weights are distributed in Safetensors and GGUF formats. However, these drafter checkpoints are not yet available through any hosted inference providers on platforms like Hugging Face. To run these models, users need to utilize a build of SGLang or llama.cpp that specifically supports DSpark for LFM2 targets. The LFM Open License v1.0 governs the commercial use of these models, permitting free commercial use for entities with annual revenues under $10 million. This license covers individual developers, startups, and small to medium-sized businesses. Larger enterprises seeking to use these models commercially must obtain a separate license directly from Liquid AI.

The release targets a variety of industries and applications where efficient, local AI processing is beneficial. This includes developer tooling, consumer applications designed to run locally on user devices, robotics and embedded systems, and sensitive workloads in healthcare, finance, and defense that require data to remain on-premise or on-device. Specific applications envisioned by Liquid AI include local coding assistants, on-device agents that perform reasoning before executing tool calls, single-user chat applications where batch sizes are minimal (batch size of 1), and offline copilots for laptop-class hardware.

Drafters, as implemented by Liquid AI, leverage speculative decoding, a technique where a smaller, faster model proposes potential outputs that a larger, more accurate model then validates. Each LFM2.5 drafter model is approximately 300 million parameters in size. Specifically, the drafter for the 1.2B-Instruct target model contains 295.7 million parameters, while the drafters for the 2.6B and 8B-A1B target models each contain 327.7 million parameters. The underlying architecture of these drafters consists of five full-attention layers. These layers are configured with a hidden size of 2048 and an intermediate size of 6144. The models employ Grouped-Query Attention (GQA) with 32 heads over 8 KV heads, and operate with a block size of 9 tokens. Notably, the drafter component itself does not ship with vocabulary weights, as these are handled by the target model.