Liquid AI’s DSpark Unleashes Faster Vision-Language Models with Up to 3.13x Speed Boost

Unleashing Speed: Liquid AI Introduces LFM2.5-VL-3B-DSpark for Faster Vision-Language Models

The bigger takeaway is simple: In the rapidly evolving world of artificial intelligence, efficiency is paramount. Vision-Language Models (VLMs), which excel at understanding and generating content based on both images and text, often face performance bottlenecks during the inference process. Liquid AI is tackling this challenge head-on with their latest innovation: LFM2.5-VL-3B-DSpark.

Meanwhile, This experimental release introduces a sophisticated speculative decoding mechanism designed to dramatically accelerate the speed of their LFM2.5-VL-3B model, promising significant performance gains without compromising the integrity or quality of the output. Let’s delve into what makes DSpark a game-changer for VLM applications.

What is LFM2.5-VL-3B-DSpark?

At its core, LFM2.5-VL-3B-DSpark is an enhancement for Liquid AI’s existing LFM2.5-VL-3B vision-language model. It incorporates a specialized “drafter” model, adding approximately 280 million parameters. The primary goal of this drafter is to predict future tokens more efficiently, thereby speeding up the decoding process for the main model.

In practical terms, Crucially, this acceleration doesn’t alter the final output of the model. This means developers and researchers can enjoy faster inference times while maintaining the same high-quality results their applications depend on.

The Magic Behind Speculative Decoding

Traditional language models generate one token at a time, requiring a separate computation (forward pass) for each. Speculative decoding introduces a clever shortcut:

• A smaller, faster drafter model proposes a sequence of several tokens ahead.
• The larger, more accurate target model then evaluates this entire block of proposed tokens in a single pass.
• It accepts the tokens it agrees with, discarding any that don’t match its own prediction.

For example, This process significantly reduces the number of full forward passes the main model needs to perform, leading to substantial speed improvements. Liquid AI’s DSpark specifically reuses the inference algorithm from their text-model DSpark drafters, demonstrating that the underlying modality (text or image) doesn’t fundamentally change how the drafter operates at the hidden layer level.

Impressive Performance Benchmarks

Liquid AI’s team reports remarkable speedups across various hardware configurations:

• Apple Silicon (M5 Max MacBook Pro): Decoding speed increased by 2.30x to an impressive 3.13x.
• NVIDIA H100 GPU: Decoding speed saw improvements ranging from 2.04x to 2.66x.
• Apple M3 Ultra: Achieved decoding speedups of 1.57x to 2.14x.

That said, These figures were derived from evaluations using the MMSpec benchmark, covering a diverse set of 6 task types, including General VQA, Image Captioning, and Complex Reasoning. It’s worth noting that the “up to” figures often come from specific tasks, such as COCO captioning yielding the 3.13x decoding speedup on the M5 Max.

Unaltered Output Quality and Temperature Considerations

A significant advantage of DSpark is its commitment to output fidelity:

• Under greedy decoding (where the model always picks the most probable next token), the output from DSpark is identical to that of the base LFM2.5-VL-3B model.
• When using non-zero temperatures for more varied outputs, speculative decoding is proven to preserve the target model’s output distribution.

However, higher temperatures can lead to a slight reduction in speed. This is because a wider probability distribution means the drafter and target model might disagree more often, resulting in fewer accepted tokens per pass and thus, lower throughput.

Deployment and Accessibility

Liquid AI has made LFM2.5-VL-3B-DSpark readily available for experimentation and deployment. The model weights are live on Hugging Face, offered in both Safetensors and GGUF formats. Furthermore, it boasts day-one support in popular frameworks:

• SGLang (v0.5.19 or newer)
• MLX-VLM (v0.7.2 or newer)
• llama.cpp

However, The release operates under the LFM Open License v1.0, which permits free commercial use for companies with under $10 million in annual revenue. This makes it accessible for many startups and smaller businesses looking to leverage advanced VLM capabilities.

Understanding End-to-End Performance on Edge Devices

While decoding speeds are impressive, it’s important to understand the concept of “end-to-end” speedup, especially on edge devices. Speculative decoding primarily accelerates the token generation phase. Other parts of the VLM pipeline, such as initial image encoding and prefill (processing visual tokens and the initial prompt), remain unaccelerated.

Meanwhile, On devices with less computational power, these unaccelerated steps constitute a larger proportion of the total latency. Liquid AI explains this through Amdahl’s Law: the overall speedup is limited by the portion of the task that isn’t optimized. For instance, on an M5 Max, a 2.69x faster decoding might translate to a 1.56x end-to-end speedup for tasks like TextVQA.

Getting Started with DSpark

Integrating DSpark into your workflow is straightforward:

• SGLang: Use v0.5.19+ and specify –speculative-algorithm DSPARK and –speculative-draft-model-path.
• MLX-VLM: Requires v0.7.2+ and accepts the drafter via –draft-model. Note that MLX-VLM currently supports greedy sampling only, so set temperature to 0.
• llama.cpp: Pair the GGUF drafter with the LFM2.5-VL-3B-GGUF target model.

In practical terms, Public integration work can be found in the respective pull requests for llama.cpp, SGLang, and MLX-VLM. Notably this initial release focuses on 16-bit weights, and acceleration for quantized models is not yet within its scope.

Key Takeaways for Developers and Researchers

• Liquid AI’s DSpark offers a significant performance boost for Vision-Language Models.
• A 279.5M parameter drafter is added, increasing the total parameter count by about 8.9%.
• Decoding speeds are up to 3.13x faster on Apple Silicon and 2.66x on an NVIDIA H100.
• The model maintains identical output under greedy decoding and preserves output distribution with sampling.
• End-to-end gains are influenced by unaccelerated phases like prefill and vision encoding, particularly on edge devices.
• Currently supports 16-bit weights; quantized model acceleration is a future consideration.

Liquid AI’s LFM2.5-VL-3B-DSpark represents a compelling step forward in making powerful Vision-Language Models more efficient and accessible. By leveraging speculative decoding, they’re paving the way for faster, more responsive AI applications across diverse hardware environments.

Expert Perspective

From an industry angle, the clearest signal around Liquid AI DSpark is how it may influence model. The story reads less like a one-day spike and more like a marker of broader movement.

The next phase will depend on how quickly teams, regulators, or customers react. In practice, that gives Liquid AI DSpark room to reshape expectations across decoding over the near term.

For readers focused on practical impact, the best next step is to watch what changes around dspark once attention turns into execution.

Frequently Asked Questions

Why does Liquid AI DSpark matter right now?

Unleashing Speed: Liquid AI Introduces LFM2.5-VL-3B-DSpark for Faster Vision-Language Models The bigger takeaway is simple: In the rapidly evolving world of artificial intelligence, efficiency is paramount.

What broader change could Liquid AI DSpark signal?

Vision-Language Models (VLMs), which excel at understanding and generating content based on both images and text, often face performance bottlenecks during the inference process.

What should the market watch next around Liquid AI DSpark?

Liquid AI is tackling this challenge head-on with their latest innovation: LFM2.5-VL-3B-DSpark.

Source: https://www.marktechpost.com/2026/09/25/liquid-ai-releases-lfm2-5-vl-3b-dspark-speculative-decoding-for-vision-language-models-with-up-to-3-13x-faster-decoding/