Liquid AI Released LFM2.5-VL-3B-DSpark Model
The new vision-language model utilizes speculative decoding to boost inference speeds on Apple and NVIDIA hardware.
Updated on Sept. 25, 2026 in Artificial Intelligence
Live Poll
Should open-source AI models require commercial licenses for companies earning over ten million dollars?
Liquid AI has launched its LFM2.5-VL-3B-DSpark model, which is designed to accelerate vision-language tasks through advanced speculative decoding. This release provides significant performance gains while maintaining open accessibility for smaller commercial entities.
Why it matters
By treating text and image tokens as identical tensors, the model allows for more efficient inference across different hardware architectures. This approach enables faster processing speeds without requiring fundamentally new algorithms for multimodal data.
The drafter model incorporates 280 million parameters across 4 attention layers using a block size of 9. It provides speed increases of up to 3.13x on Apple silicon and 2.66x on NVIDIA H100 hardware.
The players
Liquid AI
This is an artificial intelligence research organization focused on developing efficient neural network architectures.
Hugging Face
This is the primary open-source community platform used for distributing machine learning models and datasets.
The details
The model functions by having a drafter read hidden states from the target model to predict upcoming tokens, which the target then verifies in a single pass. Developers can access the weights on Hugging Face in both Safetensors and GGUF formats under the LFM Open License v1.0.
Timeline
September 25, 2026: Liquid AI officially released the LFM2.5-VL-3B-DSpark model.
The Tech Race
This release follows the trend of applying speculative decoding algorithms to multimodal vision-language models. By streamlining how tokens are proposed and verified, this technology aims to replace slower sequential inference methods that currently bottleneck many AI applications.
Developers and companies with under $10 million in annual revenue can use this model commercially for free under the LFM Open License. The improved decoding speeds mean that users can expect more responsive vision-language applications on compatible hardware.
The takeaway
This model demonstrates that optimizing hidden state processing can provide substantial performance improvements for complex vision-language tasks. Organizations should review the LFM Open License v1.0 terms to determine if their annual revenue qualifies them for free commercial deployment.
Further reading
For broader trends in machine learning efficiency, browse our Artificial Intelligence section.
Live Poll
Should open-source AI models require commercial licenses for companies earning over ten million dollars?
