Liquid AI's vision model gets a speed boost of up to 3.13 times with no change to what it says
A 280-million-parameter add-on lets LFM2.5-VL-3B generate text up to 3.13 times faster on Apple silicon, leaving output word-for-word identical.
Key points
- Liquid AI released LFM2.5-VL-DSpark in September 2026, a 280-million-parameter add-on that speeds up its LFM2.5-VL-3B vision-language model by up to 3.13 times on Apple M5 Max hardware.
- The add-on increases the base model's total size by only 8.9%.
- End-to-end response time improves by up to 2.62 times on device and up to 2.27 times on an NVIDIA H100 server chip.
- Day-one support covers llama.cpp, MLX-VLM, and SGLang, three widely used open-source inference platforms.
- Output is mathematically identical to running the base model alone, so no accuracy is sacrificed.
A vision-language model, or VLM, is an AI that can look at an image and answer questions about it in plain text. Liquid AI's LFM2.5-VL-3B is compact enough to run on a laptop rather than a data centre. The new DSpark add-on uses speculative decoding: a small, fast drafter model guesses several words ahead, and the main model checks those guesses in one batch instead of generating every word from scratch. Correct guesses skip a lot of repetitive work. Wrong ones get corrected before any bad output reaches the user.
This is the second time we've covered Liquid AI's DSpark work in a month. Our 20 August story covered the text-only drafter models; this release extends the same technique to vision.
How fast is it, really?
On an Apple M5 Max, the decoding stage runs 2.30 to 3.13 times faster by task. Full responses, including the time spent reading the image, arrive 1.56 to 2.62 times sooner. On an NVIDIA H100, the chip used in most AI server farms, the same drafter delivers end-to-end gains of 1.64 to 2.27 times.
| Hardware | Peak decoding speedup | End-to-end speedup |
|---|---|---|
| Apple M5 Max (MLX) | 3.13x | 2.62x |
| Apple M3 Ultra (llama.cpp) | 2.14x | 1.77x |
| NVIDIA H100 (SGLang) | 2.66x | 2.27x |
Those numbers come from six task types, including photo question-answering, chart analysis, and multi-turn conversation, following the MMSpec benchmark.
There's an honest ceiling. Speculative decoding only speeds up word generation. Before any words appear, the model processes the image itself, and that step stays just as slow. The more time a task spends on image processing relative to text generation, the smaller the real-world gain. Liquid AI acknowledges this directly, noting that edge devices spend a larger share of their time on image reading than server GPUs do.
For app developers, the practical implication is modest but real: responses get noticeably faster without needing a bigger server or a newer phone. The gain is sharpest on text-heavy tasks like captioning or multi-turn chat, and slimmer when the AI is mostly parsing a complex diagram.
The drafter is available on Hugging Face in Safetensors and GGUF formats with no restrictions on download or deployment.
