Liquid AI speeds up local agents with DSpark drafts for LFM2.5

Local inference is more often bottlenecked by memory bandwidth during weight transfer than by raw compute. Speculative decoding works around that problem with a draft model that proposes tokens while the target model verifies them in one pass. Liquid AI is now releasing DSpark models for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B.

Parallel drafting prunes dead ends early

DSpark combines parallel hidden-state generation with a lightweight sequential head based on a Markov chain. A verifier estimates the chance that proposed tokens will survive checking and prunes low-confidence block suffixes when verification would cost more than it saves. The draft models have roughly 300 million parameters. Under greedy decoding, the resulting sequence remains identical to the output of the target model alone.

M4 Max cuts tool-calling latency by 57 percent

The practical effect is substantial for agents running directly on a user's computer. DSpark reduced latency for LFM2.5-2.6B by an average of 57 percent across multi-tool scenarios. On an M4 Max MacBook Pro, average throughput rose from 61 to 139 tokens per second. The 1.2B model reached an average of 350 tokens per second.

llama.cpp and SGLang have support from release

The DSpark models launched with support in llama.cpp and SGLang, so users do not have to wait for integration into popular inference engines. Checkpoints are available in Safetensors and GGUF formats. Running in either engine requires a build with the relevant DSpark support.

Metal still limits the mixture-of-experts variant

The results are not uniform across architectures. LFM2.5-8B-A1B averaged only an 18 percent speedup on the M4 Max despite its higher draft acceptance rate. Liquid AI attributes that gap to the current MoE implementation in the llama.cpp Metal backend and the extra weight traffic caused by activating more experts. Further gains therefore also depend on runtime optimization.

Lilith's verdict

When a model generates more than 130 tokens per second directly on your laptop, the argument that serious agents require a cloud API starts to fall apart. This is infrastructure for local tool use without the wait.