Liquid AI 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. Each drafter introduces a speculative decoding path to an existing target model, enabling faster inference without altering the model's output. A compact ~300M-parameter draft model proposes a block of nine candidate tokens, which the target model then verifies in a single forward pass. This design trades a modest increase in memory for a substantial decoding speedupโup to 3.18x on an H100 GPU and up to 2.87x on an M4 Max MacBook Pro. Critically, the emitted sequences remain identical to the target model's standalone output under greedy decoding, so benchmark accuracy is preserved. Both llama.cpp and SGLang offer day-one support for the new checkpoints.
Deployment: Is It Ready for Production?
Yes, provided you self-host. The weights are distributed in Safetensors and GGUF formats, and the drafter checkpoints are not currently served by any hosted inference provider on Hugging Face. To run them, you'll need an SGLang or llama.cpp build with DSpark support for LFM2 targets.
Licensing Considerations
- Company Level: The LFM Open License v1.0 permits free commercial use only if your organization's annual revenue remains below $10 million. Indie developers, startups, and SMBs qualify; larger enterprises must obtain a commercial license from Liquid AI first.
- Industries: Suitable for developer tooling, local consumer apps, robotics, embedded systems, and sectors like healthcare, finance, and defense that require on-premise or on-device data handling.
- Applications: Ideal for local coding assistants, on-device AI features, edge inference, and any scenario where latency or data privacy is critical.
via MarkTechPost