LFM2.5 DSpark Draft Model Released: Inference Speed Boosted 3.18x
Liquid AI and Hugging Face have officially released the DSpark draft model checkpoints for the LFM2.5 series, covering LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. This innovation introduces a novel speculative decoding path that significantly boosts inference throughput while preserving output quality.
Key Performance Gains and Real-World Applications
Real-world tests demonstrate DSpark’s impressive acceleration. On GPUs, throughput increases by up to 3.18x, while edge devices see improvements of up to 2.87x.
In edge agent inference, LFM2.5-2.6B reduces function call latency by an average of 57%. Tested on an M4 Max MacBook Pro, it achieves speeds up to 139 tokens per second, lowering the barrier for local agent deployment and delivering a user experience that rivals proprietary cloud models.

Understanding DSpark’s Architecture and Mechanism
Traditional LLM inference is bottlenecked by memory bandwidth, as most time is spent streaming weights from DRAM to SRAM. Speculative decoding addresses this by using a lightweight draft model to generate candidate tokens, which are then validated in a single forward pass by the target model, effectively distributing weight-loading costs.
DSpark integrates existing methods through three core components:
Parallel Backbone Network: Similar to DFlash, it generates hidden states for all draft tokens in one forward pass based on the target model’s context features.Sequential Head (Markov Head): By simulating Markov chains between adjacent tokens, it enhances dependencies and improves the acceptance rate for subsequent positions.Confidence Scheduling Validator: It predicts each token’s survival probability, automatically discarding low-confidence suffixes when validation costs outweigh savings.For training, the draft model utilizes a diverse dataset mix including SFT, chat, code, and function calls. After rigorous ablation studies, the initial version features an attention-only architecture with 5 layers and 9 blocks, keeping parameters around 300 million.

Quality Assurance and Ecosystem Integration
Due to speculative decoding, greedy decoding only accepts draft tokens that perfectly match the target model’s distribution. Rejected tokens are replaced by the target model’s output, ensuring the generated sequence matches baseline greedy decoding structure with no accuracy loss in benchmarks.
On release, DSpark achieved compatibility with mainstream inference frameworks:
SGLang: Supports execution on accelerators via dedicated integration and startup configurations.llama.cpp: Offers official build support, enabling users to load corresponding GGUF weights and draft model files through the command line.
Related article
How Unitree is Shaping the Future of Humanoid Robotics
Unitree’s New Creature of Embodied AI with ultra-wide 4D LiDAR technology for advanced real-world navigation. Credit: UnitreeWang Xingxing, CEO of Unitree, targets world model breakthroughs to help humanoid machines execute 80% of tasks when placed i
Why Abnormal AI Partnered With OpenAI on Daybreak
Michael Aiello, Head of Product for Cyber at OpenAI | Credit: Michael Aiello/LinkedInAbnormal AI joins OpenAI’s Daybreak programme, as Mike Aiello says the partnership will accelerate the adoption of practical defences for enterprisesLeading frontier
AI Agent Elements Claw Completes Superconducting Material Development
As artificial intelligence pushes the boundaries of scientific exploration, a significant milestone has been achieved. On July 3, Alibaba DAMO Academy, in partnership with Renmin University of China and the University of Chinese Academy of Sciences,
Related Special Topic Recommendations
Comments (0)
0/500
Liquid AI and Hugging Face have officially released the DSpark draft model checkpoints for the LFM2.5 series, covering LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. This innovation introduces a novel speculative decoding path that significantly boosts inference throughput while preserving output quality.
Key Performance Gains and Real-World Applications
Real-world tests demonstrate DSpark’s impressive acceleration. On GPUs, throughput increases by up to 3.18x, while edge devices see improvements of up to 2.87x.
In edge agent inference, LFM2.5-2.6B reduces function call latency by an average of 57%. Tested on an M4 Max MacBook Pro, it achieves speeds up to 139 tokens per second, lowering the barrier for local agent deployment and delivering a user experience that rivals proprietary cloud models.

Understanding DSpark’s Architecture and Mechanism
Traditional LLM inference is bottlenecked by memory bandwidth, as most time is spent streaming weights from DRAM to SRAM. Speculative decoding addresses this by using a lightweight draft model to generate candidate tokens, which are then validated in a single forward pass by the target model, effectively distributing weight-loading costs.
DSpark integrates existing methods through three core components:
Parallel Backbone Network: Similar to DFlash, it generates hidden states for all draft tokens in one forward pass based on the target model’s context features.Sequential Head (Markov Head): By simulating Markov chains between adjacent tokens, it enhances dependencies and improves the acceptance rate for subsequent positions.Confidence Scheduling Validator: It predicts each token’s survival probability, automatically discarding low-confidence suffixes when validation costs outweigh savings.For training, the draft model utilizes a diverse dataset mix including SFT, chat, code, and function calls. After rigorous ablation studies, the initial version features an attention-only architecture with 5 layers and 9 blocks, keeping parameters around 300 million.

Quality Assurance and Ecosystem Integration
Due to speculative decoding, greedy decoding only accepts draft tokens that perfectly match the target model’s distribution. Rejected tokens are replaced by the target model’s output, ensuring the generated sequence matches baseline greedy decoding structure with no accuracy loss in benchmarks.
On release, DSpark achieved compatibility with mainstream inference frameworks:
SGLang: Supports execution on accelerators via dedicated integration and startup configurations.llama.cpp: Offers official build support, enabling users to load corresponding GGUF weights and draft model files through the command line.
How Unitree is Shaping the Future of Humanoid Robotics
Unitree’s New Creature of Embodied AI with ultra-wide 4D LiDAR technology for advanced real-world navigation. Credit: UnitreeWang Xingxing, CEO of Unitree, targets world model breakthroughs to help humanoid machines execute 80% of tasks when placed i
Why Abnormal AI Partnered With OpenAI on Daybreak
Michael Aiello, Head of Product for Cyber at OpenAI | Credit: Michael Aiello/LinkedInAbnormal AI joins OpenAI’s Daybreak programme, as Mike Aiello says the partnership will accelerate the adoption of practical defences for enterprisesLeading frontier
AI Agent Elements Claw Completes Superconducting Material Development
As artificial intelligence pushes the boundaries of scientific exploration, a significant milestone has been achieved. On July 3, Alibaba DAMO Academy, in partnership with Renmin University of China and the University of Chinese Academy of Sciences,





Home






