Liquid AI has announced the release of LFM2.5-VL-DSpark, a new vision-language model that extends the company’s efficient foundation model architecture to multimodal understanding. The announcement, made on September 24, 2026, marks Liquid AI’s entry into the competitive vision-language model space with a focus on computational efficiency.
The LFM2.5-VL-DSpark model builds on the company’s Liquid Foundation Models (LFM) approach, which uses dynamically expanding computational pathways rather than the Mixture-of-Experts architecture popularised by Mistral and Meta. This architecture allows the model to allocate computational resources adaptively based on input complexity.
Efficient Multimodal Understanding
Unlike traditional vision-language models that process images and text through separate pathways, LFM2.5-VL-DSpark implements a unified approach where visual and textual information flow through the same dynamic computation pathways. This design choice reduces the overall parameter count needed while maintaining competitive performance on multimodal benchmarks.
The “DSpark” designation refers to the model’s dynamic sparse activation pattern, where only a subset of parameters engage for any given input. This approach has proven effective in Liquid AI’s text-only models, and the extension to vision-language tasks demonstrates the architecture’s flexibility.
Early benchmark results show LFM2.5-VL-DSpark achieving competitive scores on common vision-language benchmarks including MMMU, MathVista, and ChartQA, while requiring significantly fewer computational resources than comparable models from OpenAI and Google.
Enterprise Implications
The release targets enterprises seeking vision capabilities without the infrastructure costs of larger models. Liquid AI’s focus on efficiency positions the model for deployment in cost-sensitive environments where multimodal AI was previously impractical.
“Vision-language capabilities shouldn’t require frontier-scale infrastructure,” said a Liquid AI spokesperson. “LFM2.5-VL-DSpark demonstrates that efficient architectures can deliver strong multimodal performance at a fraction of the cost.”
The model is available via API with competitive pricing, and open-weight releases are planned for the fourth quarter of 2026. This strategy mirrors the company’s approach with earlier LFM releases, offering both API access and self-hosted deployment options.
Competitive Landscape
The vision-language market has grown increasingly competitive, with OpenAI’s GPT-4V, Google’s Gemini Vision, and Anthropic’s Claude Vision dominating enterprise deployments. Liquid AI enters the market with a different value proposition: efficiency over raw capability.
The company’s approach appeals to organizations prioritizing cost control and deployment flexibility over maximum performance. As AI spending scrutiny increases across enterprises, efficient alternatives may gain market share against more expensive alternatives.
LFM2.5-VL-DSpark represents Liquid AI’s bid to become the default choice for enterprises seeking capable vision-language AI without the premium pricing of frontier models.