Up to 3.2x Faster Inference with LFM2.5-DSpark

LiquidAI Releases LFM2.5-DSpark for Faster On-Device and Server Inference

LiquidAI has released 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. The new checkpoints introduce a speculative decoding path that accelerates inference speed by up to 3.2x without altering output quality. This performance boost is notable as it applies to both high-end server hardware like the NVIDIA H100 and consumer-grade systems such as an Apple M4 Max MacBook Pro, making advanced AI models more responsive for a wider range of applications.

Technical Approach and Performance Metrics

The performance gains are achieved through DSpark, a speculative decoding technique designed to overcome the memory-bound latency typical in LLM inference. Instead of generating tokens one by one, a small, ~300M parameter draft model produces a sequence of candidate tokens. The larger target model then verifies these candidates in a single, efficient forward pass. This approach is available with day-one support in widely used inference engines like llama.cpp and SGLang, facilitating immediate adoption by the developer community. Key performance uplifts include:

• Up to 3.18x throughput improvement on an NVIDIA H100 GPU.
• Up to 2.87x throughput improvement on an Apple M4 Max device.
• A 57% average reduction in function-calling latency for the LFM2.5-2.6B model, directly benefiting agentic workflows.
• Guaranteed output parity with the original models under greedy decoding.

Impact on the AI Ecosystem

By open-sourcing these DSpark-enabled models, LiquidAI provides a direct path for developers to build more interactive and performant applications. The speedup on consumer hardware is particularly significant, pushing on-device model throughput well beyond that of many proprietary cloud APIs. For example, the LFM2.5-2.6B model reaches an average of 139 tok/s on an M4 Max. While the Mixture-of-Experts model, LFM2.5-8B-A1B, sees more modest on-device gains due to current limitations in llama.cpp's Metal backend, the server-side performance still shows a substantial 2.54x average speedup, highlighting the technique's broad utility.

LiquidAI's application of speculative decoding with DSpark demonstrates that significant performance gains are still achievable at the inference level, shifting focus from pure model scale to architectural efficiency for both cloud and edge deployments.