Liquid AI Ships DSpark Checkpoints for Up to 3.18x Faster LFM2.5 Inference

Liquid AI released speculative-decoding draft checkpoints for three models in its LFM2.5 family on August 20, 2026. [1] The checkpoints cover LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and the LFM2.5-8B-A1B mixture-of-experts model, each adding roughly 300 million parameters of draft overhead. [1] A draft proposes a block of nine candidate tokens, and the target model verifies the whole block in a single forward pass. [2]

Throughput on a single H100 GPU rose up to 3.18x, while an Apple M4 Max MacBook Pro saw gains up to 2.87x. [1][2] Because the target model only emits verified tokens, the generated text is identical under greedy decoding, leaving benchmark accuracy unchanged. [1][2]

The weights ship in Safetensors and GGUF formats with day-one support in llama.cpp and SGLang. [1][2] The drafter checkpoints are deployable via self-hosting, as they are not served by hosted inference providers on Hugging Face today. [2] The company's license allows free commercial use for entities under $10 million in annual revenue, but larger enterprises must contact Liquid AI. [2]

The implementation of DSpark—introduced in a July 2026 paper by DeepSeek researchers—addresses the memory-bound decode phase of LLM inference. By lowering inference latency, these models specifically target on-device agentic workloads, cutting multi-tool function-calling latency by 57% on average for the 2.6B model. This pushes capable local agents closer to viability on consumer hardware like the MacBook Pro. What the sources don't address: whether hosted inference providers will eventually offer DSpark-enabled endpoints for developers who prefer not to self-host.