Artificial intelligence models are becoming increasingly capable, but faster inference remains an important challenge, particularly for applications that need low latency. One approach attracting attention is speculative decoding, a technique that uses a smaller draft model to predict upcoming tokens while a larger target model verifies those predictions.
Recent community releases have brought DSpark speculative-decoding draft models to Liquid AI’s LFM2.5 family. These models are designed to work with LFM2.5 target models through inference frameworks such as SGLang, providing developers with an experimental approach to accelerating token generation.
What Is LFM2.5?
Liquid AI’s LFM2.5 is a family of hybrid foundation models designed with efficient, on-device deployment in mind. The official LFM2.5 collection includes models such as LFM2.5-230M, LFM2.5-350M, LFM2.5-2.6B and LFM2.5-8B-A1B, along with additional task-specific variants.
The models are aimed at applications where efficiency, memory usage and local inference are important. Liquid AI has positioned the LFM2.5 family for deployment across different hardware environments rather than relying exclusively on large data-center models.
What Is DSpark?
DSpark is a speculative-decoding approach in which a smaller draft model proposes several future tokens and the target model verifies those predictions.
For example, a verified community DSpark draft for LFM2.5-8B-A1B is designed specifically as a speculative-decoding draft model for the LiquidAI/LFM2.5-8B-A1B target. Its model card describes a lightweight block drafter with a confidence head and a low-rank Markov transition head.
A similar DSpark draft has been released for LFM2.5-1.2B-Instruct. The available versions include different numbers of draft layers, including 2-layer, 3-layer and 5-layer configurations.
How Speculative Decoding Works
Traditional autoregressive language-model generation generally produces tokens sequentially. Speculative decoding attempts to reduce this bottleneck.
The process can be understood in three stages:
1. A Smaller Model Makes Predictions
The draft model quickly predicts a short sequence of tokens that it believes the larger target model will generate.
2. The Target Model Verifies the Predictions
Instead of generating every token independently, the larger LFM2.5 target model evaluates the proposed sequence.
3. Correct Predictions Are Accepted
When the proposed tokens satisfy the target model’s verification process, multiple tokens can potentially be accepted in a single decoding step. Incorrect predictions can be rejected and regenerated by the target model.
This approach can improve generation throughput when the draft model is sufficiently fast and its predictions align well with the target model.
DSpark and LFM2.5 Integration
The available DSpark model cards show that these drafts are specifically configured to work with LFM2.5 target models.
For example, the LFM2.5-8B-A1B DSpark draft can be launched with SGLang using the DSPARK speculative algorithm, the LFM2.5-8B-A1B target model and the corresponding DSpark draft model.
The LFM2.5-1.2B-Instruct DSpark models similarly provide SGLang configuration for speculative decoding.
This makes the development particularly interesting for developers experimenting with local or self-hosted LFM2.5 inference.
Why Faster Decoding Matters
Reducing inference latency can be valuable in applications such as AI assistants, coding tools, local agents and interactive applications.
Higher decoding efficiency can potentially mean:
- Faster response generation
- Better interactive user experiences
- Greater throughput from available hardware
- More efficient use of compute resources
- Improved practicality of local AI deployment
However, the actual speed improvement depends on hardware, software implementation, target model, draft model, workload and acceptance rate. Therefore, benchmark results should be evaluated under clearly defined test conditions rather than assuming a particular speed increase across all systems.
Different DSpark Draft Configurations
One interesting aspect of the current releases is that DSpark is not represented by a single universal draft model.
For LFM2.5-1.2B-Instruct, for example, available community drafts include different configurations such as 2-layer, 3-layer and 5-layer versions. The 3-layer configuration uses a Qwen3-style GQA block drafter, a rank-256 Markov transition head and a confidence head.
The 8B-A1B target also has a DSpark draft configuration, with the corresponding model card describing a 2-layer drafter and a seven-token block structure consisting of an anchor plus six draft positions.
An Important Point About the 3.18× Claim
Claims of a specific 3.18× decoding improvement should not be treated as a universal performance guarantee.
The sources currently available for the DSpark LFM2.5 drafts confirm the speculative-decoding implementation and integration with LFM2.5 models, but I would not state that every LFM2.5 deployment becomes 3.18× faster without a benchmark that specifies the hardware, software stack, workload and comparison methodology.
This distinction is important because speculative decoding performance can vary substantially depending on how frequently the draft model’s predictions are accepted by the target model.
What This Means for Local AI
The combination of efficient LFM2.5 models and speculative decoding is particularly relevant to the growing local-AI ecosystem.
Liquid AI’s LFM2.5 family already focuses on efficient deployment, with models ranging from hundreds of millions of parameters to larger models such as LFM2.5-8B-A1B.
Adding speculative-decoding techniques could provide another way for developers to improve the responsiveness of these models without simply increasing hardware requirements.
The Bigger Picture
The development highlights an important direction in AI inference: better efficiency does not always require a larger model or more powerful hardware.
Techniques such as speculative decoding attempt to make existing models work more efficiently by changing how token generation is performed. For developers running AI locally or on constrained infrastructure, improvements in decoding efficiency can be particularly valuable.
At the same time, DSpark should currently be viewed as an experimental/community implementation around Liquid AI’s LFM2.5 models, rather than describing it as a formally announced new Liquid AI model family unless Liquid AI itself makes such an announcement.
Conclusion
LFM2.5 is part of Liquid AI’s push toward efficient foundation models suitable for on-device and local deployment. Recent DSpark draft models demonstrate how speculative decoding can be paired with LFM2.5 targets to explore faster inference through draft-and-verify generation.
The technology is particularly interesting for developers building local AI assistants, agents and other latency-sensitive applications. However, performance should be judged using reproducible benchmarks rather than relying on a single headline speed figure.
The broader trend is clear: as AI models become more capable, improving inference efficiency, latency and hardware utilisation will be just as important as increasing model size and capability.
Disclaimer
This article is based on publicly available Liquid AI model information and community-published DSpark model documentation. DSpark performance can vary depending on the model, hardware, inference framework and workload. Specific speed improvements should be independently benchmarked before being used as performance guarantees.