What was released
Liquid AI published LFM2.5-DSpark draft checkpoints on August 20 for its 1.2B, 2.6B and 8B target models. The draft models are around 300 million parameters and are intended to accelerate generation through speculative decoding.
The release includes support in llama.cpp and SGLang, which makes the technique accessible across local devices and GPU serving stacks. The model cards provide the artifacts teams need to test rather than only a performance announcement.
How speculative decoding works
A small draft model proposes several likely next tokens. The larger target model checks those candidates in a more parallel operation and accepts the valid prefix. When the draft is accurate, the system produces multiple target-quality tokens with fewer sequential target-model steps.
The output distribution is still governed by the target model when the algorithm is implemented correctly. The performance gain depends on how often the draft and target agree, the cost of checking candidates and the overhead of running both models.
The reported results
Liquid AI reports up to 3.18 times higher throughput on GPU and up to 2.87 times on-device. For the 2.6B configuration, the company reports an average 57 percent reduction in function-calling latency across its tests.
Those are vendor results, not independent guarantees. The word 'up to' also describes a best observed condition. Prompt length, batch size, generation length, quantization, device memory and task type can move the result substantially.
Why function calling is a useful target
Agent workflows frequently wait for short structured outputs before they can call a tool. Lowering that latency can improve the time to first action even when a long-form text benchmark would show a different advantage.
However, speed is not enough. Teams should validate JSON correctness, tool selection, argument accuracy and recovery after a rejected call. A faster invalid request can increase operational cost rather than reduce it.
A fair local test
Benchmark the target model alone and the speculative pair with identical prompts, sampling settings and output constraints. Record time to first token, inter-token latency, total completion time, memory use, energy where measurable and exact output validity.
Run representative workloads instead of a single synthetic prompt. The release is valuable because it gives local AI teams another optimization path; the decision to adopt it should still rest on repeatable measurements from the device and application that will carry the production load.
Sources & further reading
Social-media activity is treated as a signal of attention, not proof. Product claims are attributed to the linked publisher or announcement.