Accelerating vision-language models with LFM2.5-VL-DSpark

The world of Artificial Intelligence is moving at an incredible pace, and one area seeing rapid development is Vision-Language Models (VLMs). These powerful models can understand and generate content from both images and text, opening doors to a new generation of intelligent applications. However, with great power often comes significant computational cost, especially during the inference phase – the process where the model generates its outputs. This is where innovations like LFM2.5-VL-DSpark step in, promising to make these sophisticated models run much faster and more efficiently.

Liquid AI, a company focused on developing open-weight foundation models, has introduced LFM2.5-VL-DSpark as a crucial accelerator for its LFM2.5-VL-3B vision-language model. This isn't a new VLM in itself, but rather a specialized "draft model" built using the DSpark speculative decoding framework. Its sole purpose is to supercharge the inference speed of its target VLM, ensuring that AI practitioners and developers can deploy high-quality multimodal AI with significantly reduced latency.

What Exactly Is LFM2.5-VL-DSpark?

To understand LFM2.5-VL-DSpark, we first need to briefly touch upon Vision-Language Models (VLMs) and the challenge of their inference speed. VLMs are neural networks capable of processing and relating information from both visual inputs (like images or videos) and textual inputs. They can answer questions about images, generate descriptions, perform visual grounding (identifying objects based on text queries), and much more. Think of applications like advanced image search, AI assistants that can "see" and understand your screen, or tools that automatically generate detailed alt-text for accessibility.

The problem is that running these large, complex models, especially when generating long or detailed responses (the "decode phase"), can be slow and resource-intensive. This slowness impacts user experience, particularly in interactive applications or edge deployments where quick responses are critical. Traditional inference involves the model generating tokens one by one, which is computationally heavy.

LFM2.5-VL-DSpark is Liquid AI's answer to this bottleneck for their LFM2.5-VL-3B model. It's a "speculative decoding draft model." In simpler terms, it's a smaller, faster model (the "drafter") that works in tandem with the larger, more powerful LFM2.5-VL-3B (the "target model"). The drafter's job is to quickly predict a sequence of upcoming tokens, which the target model then efficiently verifies in a single pass. This dramatically speeds up the overall process without compromising the quality of the output.

The weights for LFM2.5-VL-3B-DSpark were released on Hugging Face on September 18, 2026, with Liquid AI's official announcement following on September 24, 2026. The underlying DSpark technique itself originates from research by DeepSeek.

Why Does Speculative Decoding with DSpark Matter?

The significance of LFM2.5-VL-DSpark lies in its ability to bring high-performance VLM capabilities closer to practical, real-world applications. Here's why this acceleration technology is a big deal for AI practitioners:

• Faster User Experiences: For interactive AI applications, latency is everything. Whether it's an AI agent describing an image in real-time or a multimodal chatbot, faster responses lead to a smoother, more natural user experience. LFM2.5-VL-DSpark directly targets this, making VLM interactions feel more immediate.
• Efficient Resource Utilization: By speeding up inference, DSpark allows developers to get more work done with the same hardware. This can translate to lower operational costs in cloud deployments or enable more sophisticated AI to run effectively on resource-constrained edge devices, such as laptops or mobile phones.
• Maintained Output Quality: A critical advantage of speculative decoding, and DSpark in particular, is that it guarantees the output quality remains identical to what the larger, unaccelerated target model would produce under greedy decoding. The drafter only proposes tokens; the target model's verification ensures correctness. This means developers don't have to trade off speed for accuracy.
• Broader Deployment Possibilities: The ability to run complex VLMs faster opens up new avenues for deployment. Imagine advanced visual AI capabilities integrated directly into local applications, offering privacy and offline functionality that cloud-based solutions can't.

How Does LFM2.5-VL-DSpark Work Its Magic?

The core principle behind LFM2.5-VL-DSpark is speculative decoding, a technique designed to optimize the token generation process in large language models. Here's a high-level breakdown of how it functions with vision-language models:

1. Drafting Candidate Tokens: When the VLM needs to generate the next part of a response, the smaller LFM2.5-VL-DSpark drafter model quickly generates a "draft" sequence of multiple candidate tokens in a single forward pass. This is much faster than the target model generating them one by one.
2. Parallel Verification: Instead of processing tokens sequentially, the larger LFM2.5-VL-3B target model takes this entire block of drafted tokens and verifies them all at once. It essentially checks if the drafter's predictions match what it would have produced.
3. Acceptance and Correction: If the drafted tokens are correct, they are accepted, and the process continues with the next block. If the target model finds a discrepancy, it rejects the incorrect portion of the draft and generates the correct token itself, then continues drafting from that point. Because the target model always has the final say, the output is guaranteed to be the same as if the target model had run alone.

The DSpark framework itself introduces specific enhancements to this general speculative decoding approach. It incorporates three key ideas: a parallel backbone for efficient block drafting, a lightweight sequential module (modeled as a Markov chain) to improve acceptance rates at the end of a block, and a confidence-scheduled verifier that intelligently prunes low-confidence suffixes to avoid unnecessary computation.

For VLMs, the vision input (image patches) and text tokens are projected into a shared representation before these layers, allowing the drafter to operate uniformly regardless of the input modality. This means the speculative decoding mechanism works seamlessly across both visual and textual contexts.

Key Features and Technical Details

LFM2.5-VL-DSpark is engineered for efficiency and compatibility within the AI ecosystem:

• Compact Drafter Size: The LFM2.5-VL-DSpark drafter model is relatively small, adding approximately 280 million parameters. This represents only an 8.9% increase in parameters on top of the 3-billion-parameter LFM2.5-VL-3B target model, making the memory overhead minimal for the significant speed gains.
• Optimized Architecture: The drafter uses a simplified attention-only architecture with 4 layers. During training, a block size of 9 tokens was used, and for inference, a block size of 8 or 9 is recommended, with 8 specifically for Apple silicon.
• Broad Framework Support: Liquid AI has ensured "day-one support" for popular inference frameworks. This includes llama.cpp, MLX-VLM, and SGLang. These integrations are open-sourced and contributed upstream to the official codebases, making it easier for developers to incorporate DSpark into their existing workflows.
• Quality Parity: As highlighted, a cornerstone of DSpark is its guarantee that the output generated is identical to what the base model would produce without acceleration, assuming greedy decoding. This eliminates concerns about reduced accuracy or "hallucinations" often associated with speed-up techniques.

Performance and Benchmarks

Liquid AI's internal benchmarks showcase impressive speedups for LFM2.5-VL-DSpark. It's important to note these are vendor-measured figures, and independent reproductions are still emerging.

• Decode Speedups: The model achieves up to 3.13x decode speedup on devices like M-series Macs and up to 2.66x on an H100 GPU.
• End-to-End Gains: When looking at the full end-to-end inference process, the gains are also substantial: up to 2.62x on device and up to 2.27x on an H100.
• Hardware-Specific Performance:
  • On an H100 GPU across six vision tasks, decode speedups ranged from 2.04x to 2.66x, with end-to-end improvements between 1.64x and 2.27x.
  • On an M5 Max, decode speedups were observed between 2.30x and 3.13x, leading to end-to-end gains of 1.56x to 2.62x.
  • For an M3 Ultra, decode acceleration ranged from 1.57x to 2.14x, with end-to-end benefits of 1.30x to 1.77x.
• Acceptance Rate: The drafter typically sees an acceptance rate of roughly 3.2 to 4.5 tokens per target verification pass, indicating its efficiency in predicting correct token sequences.

It's crucial to understand that while these decode speedups are significant, the overall end-to-end gains are slightly lower. This is because speculative decoding primarily accelerates the token generation (decode) phase. The initial vision encoding and prefill steps, which also contribute to the total latency in VLMs, are not directly accelerated by DSpark.

Implications for AI Practitioners and Developers

For anyone working with or planning to deploy Vision-Language Models, LFM2.5-VL-DSpark offers compelling advantages:

• Enhanced Interactive AI: Developers building real-time AI agents, interactive visual assistants, or applications requiring quick visual understanding can leverage DSpark to make their systems far more responsive. This is particularly beneficial for conversational AI where human-like interaction speeds are desired.
• Edge AI Development: The on-device acceleration, especially on Apple Silicon, is a game-changer for edge AI. It means more complex VLM tasks can be performed directly on user devices, reducing reliance on cloud infrastructure, improving privacy, and enabling offline capabilities. Liquid AI has also shown their LFM2.5-VL-450M model running entirely on the Apple Neural Engine, highlighting their commitment to on-device AI.
• Cost-Effective Deployment: Faster inference means less compute time per request. For cloud deployments, this directly translates to lower costs for GPU usage. For on-premise deployments, it means existing hardware can handle higher throughput.
• Seamless Integration: With direct support for llama.cpp and SGLang, integrating LFM2.5-VL-DSpark into existing VLM pipelines is designed to be straightforward. Developers can leverage these established frameworks to quickly implement the acceleration.

While the LFM2.5-VL-3B model itself is open-weight, Liquid AI's licensing for commercial use by entities with over $10M annual revenue may require a separate commercial license. For individual developers or smaller businesses, the open-weight nature allows for extensive experimentation and deployment.

Conclusion

LFM2.5-VL-DSpark represents a significant step forward in making powerful Vision-Language Models more practical and efficient. By intelligently accelerating the inference process through speculative decoding, Liquid AI is enabling developers to build faster, more responsive, and more cost-effective multimodal AI applications without sacrificing output quality. As AI continues to integrate into our daily lives and tools, innovations like DSpark are essential for pushing the boundaries of what's possible, especially in interactive and on-device scenarios.

Frequently Asked Questions

What is LFM2.5-VL-DSpark?

LFM2.5-VL-DSpark is a speculative decoding draft model developed by Liquid AI. It is designed to accelerate the inference speed of Liquid AI's LFM2.5-VL-3B vision-language model, rather than being a standalone VLM itself.

How does LFM2.5-VL-DSpark improve VLM inference speed?

It uses a technique called speculative decoding. A smaller, faster "drafter" model (LFM2.5-VL-DSpark) quickly generates a block of candidate tokens, which the larger LFM2.5-VL-3B model then verifies all at once. This parallel verification is significantly faster than the traditional token-by-token generation.

Does LFM2.5-VL-DSpark affect the quality of the VLM's output?

No, under greedy decoding, LFM2.5-VL-DSpark guarantees that the output quality is identical to what the LFM2.5-VL-3B model would produce on its own. The larger target model always verifies the proposed tokens, ensuring correctness.

What kind of speedups can I expect with LFM2.5-VL-DSpark?

Liquid AI reports decode speedups of up to 3.13x on devices like M-series Macs and 2.66x on H100 GPUs. End-to-end gains, considering the entire inference process, can reach up to 2.62x on device and 2.27x on H100.