Speeding Up Vision-Language Workloads Using LFM2.5-VL-DSpark

🔍 Read the full analysis: Speeding Up Vision-Language Workloads Using LFM2.5-VL-DSpark on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Liquid AI has released LFM2.5-VL-DSpark, an experimental 280M-parameter draft model that accelerates inference of its open-weight LFM2.5-VL-3B vision-language model through speculative decoding. According to the company, the drafter adds just 8.9% to the target model’s parameter count while delivering decoding speedups of up to 3.13x on Apple silicon and up to 2.66x on an NVIDIA H100 — without changing output quality. The model is available now on Hugging Face in Safetensors and GGUF formats, with day-one support in llama.cpp, MLX-VLM, and SGLang.

The drafter extends Liquid AI’s DSpark recipe — previously applied to its text-only LFM2.5 models — to a multimodal target for the first time. It captures the target model’s hidden states at a fixed set of tapped layers and drafts blocks of candidate tokens conditioned on those states. Because image patches and text tokens are projected into a shared representation before the tapped layers, the drafter operates on hidden-state vectors of identical dimensionality regardless of input modality, and the inference algorithm is unchanged from the text models.

The final drafter is a simplified attention-only model with 4 layers, selected via ablations across 3, 4, and 5 layers, with a block size of 9. Liquid AI reports that acceptance improved over 10 training epochs before reaching diminishing returns. The drafter comprises a 193.0M-parameter decoder stack, a 21.0M hidden-state projection, a 65.5M Markov head, and roughly 6.4k parameters in norms and a confidence head — about 279.5M in total. It was trained on a mixture of vision-language supervised fine-tuning data weighted toward expected serving workloads.

Benchmarks follow the MMSpec protocol across six vision tasks: general VQA, text VQA, image captioning, chart VQA, complex reasoning, and multi-turn conversation. On-device, with MLX on an M5 Max, decoding runs 2.30x to 3.13x faster and end-to-end latency improves 1.56x to 2.62x. With llama.cpp on an M3 Ultra, decoding improves 1.57x to 2.14x, end-to-end 1.30x to 1.77x. On H100, the company reports decoding speedups ranging up to 2.66x and end-to-end gains of 1.64x to 2.27x, using a DSpark block size of 8.

Making Multimodal AI Practical on Laptops

The release targets a persistent bottleneck for local and edge AI: vision-language models are slower than text models because images must pass through a vision encoder and then be processed as hundreds of visual tokens alongside the text prompt. A drafter that roughly doubles or triples decode speed — for under 9% more parameters — could make 3B-class multimodal models practical on laptops and phones, where users feel latency directly.

Day-one integration matters as much as the speedup itself. DSpark support in llama.cpp, MLX-VLM, and SGLang means the acceleration works in the toolchains hobbyists and deployers already use, rather than requiring custom inference code. Combined with Liquid AI’s open-weight licensing — download, fine-tune, and deploy without restrictions, per the company — the release positions the LFM2.5 family for on-device multimodal applications.

How Speculative Decoding Keeps Outputs Identical

Speculative decoding is an established technique in which a small, fast “draft” model proposes candidate tokens that the larger target model verifies in batch, accepting matching tokens and discarding the rest. Because verification is cheaper than sequential generation, accepted drafts translate into net speedup with identical outputs under greedy decoding. The technique is exact: the target model verifies every proposed token, so greedy output matches the target alone.

Liquid AI released its first LFM2.5-DSpark drafter models for text-only LFM2.5 targets earlier in 2026. The new vision drafter reuses the same architecture and inference algorithm, differing mainly in training data and the shared multimodal representation. The LFM2.5 family spans base models, audio, and vision variants, with the 3B vision model positioned as an edge-capable multimodal option.

“It adds a speculative decoding path that trades a minimal increase in memory footprint for a larger speedup without changing output quality.”

— Liquid AI, announcement post

Experimental Status and Benchmark Caveats

The release is explicitly labeled experimental, and the company has not indicated when or whether the drafter will be promoted to a stable release. The reported speedups are Liquid AI’s own measurements, not independently verified third-party benchmarks, and results will vary with hardware, prompt composition, and image resolution.

The company is candid about the ceiling: speculative decoding accelerates only the decode phase, not vision encoding or prefill. On edge devices with limited compute, those stages consume a larger share of wall-clock time, so end-to-end gains (1.30x–2.62x) trail decode gains — an illustration of Amdahl’s law that Liquid AI itself highlights.

One figure in the company’s GPU results appears internally inconsistent — a lower bound described as “20.4x” in a range stated as “20.4x to 2.66x” — which reads as a typo, likely for 2.04x; Liquid AI has not clarified the figure. It is also unclear how acceptance rates behave on out-of-distribution vision tasks, how the drafter affects sampling-based (non-greedy) generation quality, and what the memory footprint increase is in runtime terms beyond parameter count. Details of the training data mixture have not been published.

Community Benchmarks and Stable Release Path

The model and required integration patches are public: SGLang support requires a build with DSpark for LFM2.5 targets (PR #40651), llama.cpp requires PR #29339, and MLX-VLM requires PR #2280. Early adopters will likely publish independent benchmarks, which will test whether Liquid AI’s reported speedups hold across diverse hardware and workloads.

Open questions include whether Liquid AI promotes the drafter from experimental to stable status, whether the DSpark recipe extends to other vision-language model sizes or modalities such as audio, and whether the company publishes details of the training data mixture and clarifies the inconsistent GPU benchmark figure.

Key Questions

What is LFM2.5-VL-DSpark?

It is an experimental 280M-parameter draft model from Liquid AI that accelerates the company’s LFM2.5-VL-3B vision-language model using speculative decoding. It proposes candidate tokens that the larger model verifies, speeding up generation without changing outputs under greedy decoding.

How much faster is inference with the drafter?

According to Liquid AI’s own benchmarks, decoding runs up to 3.13x faster on Apple silicon (M5 Max with MLX) and up to 2.66x faster on an NVIDIA H100. End-to-end latency gains are lower, ranging from 1.30x to 2.62x, because speculative decoding does not accelerate vision encoding or prefill.

Does using the drafter change the model’s outputs?

No, under greedy decoding. Speculative decoding is exact: the target model verifies every proposed token, so outputs match those of the target model running alone. How the drafter affects sampling-based (non-greedy) generation quality is not yet documented.

Where can I use LFM2.5-VL-DSpark?

The model is available on Hugging Face in Safetensors and GGUF formats, with day-one support in llama.cpp, MLX-VLM, and SGLang. Each requires a specific merge or build with DSpark support (llama.cpp PR #29339, MLX-VLM PR #2280, SGLang PR #40651).

What are the limitations of the release?

The drafter is labeled experimental, the speedups are company-reported rather than independently verified, and results vary by hardware and workload. It only accelerates the decode phase, and details of the training data mixture and runtime memory footprint beyond parameter count have not been published.

Primary source: Hugging Face · via ThorstenMeyerAI.com