Liquid AI Launches LFM2.5-VL-DSpark: 3.13x Decoding Speedup at 8.9% Memory Cost
Liquid AI has released LFM2.5-VL-DSpark, a speculative draft model for its LFM2.5-VL-3B vision-language model, delivering up to 3.13x faster decoding on edge devices and 2.66x on H100 GPUs with only an 8.9% parameter increase (280M extra parameters). The model supports llama.cpp, MLX-VLM, and SGLang from day one, and is distributed in Safetensors and GGUF formats on Hugging Face under an open-weights license.
Executive Overview
Liquid AI has officially announced in its blog (September 2026) the release of the experimental draft model LFM2.5-VL-DSpark, a companion to its LFM2.5-VL-3B vision-language model. The model is immediately available on Hugging Face in both Safetensors and GGUF formats under an open-weights license. It achieves up to 3.13x decoding speedup on edge devices and 2.66x on an H100 GPU, with only an 8.9% increase in parameter count (280M additional parameters). This release marks a significant step in making high-performance multimodal inference accessible on consumer hardware and budget-constrained servers.
📊 Official Data & Technical Specifications Sheet
| Technical Axis | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Open-weights and free to download from Hugging Face with no licensing fees; no subscription tiers or token prices announced because distribution is via weights, not a paid API. |
| 🌐 Platforms & Immediate Availability | Hugging Face (Safetensors + GGUF); day-one support for llama.cpp (PR#29339), MLX-VLM (PR#2280), and SGLang (PR #40651) with an OpenAI-compatible endpoint at http://localhost:30000/v1. |
| ⚡ Performance & Speed Benchmarks | Decoding speedup: up to 3.13x on-device (MLX/M5 Max) and 2.66x on H100. End-to-end latency: up to 2.62x on-device and 2.27x on H100. On llama.cpp/M3 Ultra: 1.57x–2.14x decoding and 1.30x–1.77x end-to-end. |
| 🛡️ Security & Tamper Resistance | Exact speculative decoding: the target model verifies every proposed token, so greedy outputs are identical to the target model alone without distortion. No separate security benchmarks announced in the release. |
| 🧠 Context Window | No separate numeric context window size announced; the model operates as a draft on the hidden states of the target LFM2.5-VL-3B with a block size of 8 or 9 tokens. |
| 🌍 Arabic & Regional Support | No separate Arabic evaluation in the official release; the model is multimodal (text + image) and globally available via Hugging Face without geographic restrictions. |
Deep-Dive Features & Architecture
LFM2.5-VL-DSpark builds on the same architecture as the LFM2.5-DSpark text draft models: it captures the hidden states of the target model at a fixed set of tapped layers, then generates a block of k candidate tokens. Image patches and text tokens are projected into a shared representation before those layers, so the model operates on vectors of the same dimension regardless of input type—meaning the inference algorithm is unchanged from the text-only draft models. The resulting model is a simplified attention-only draft with 4 layers and a block size of 9, trained for 10 epochs on a mix of vision-language SFT data, with acceptance rate improving as training tokens increase until diminishing returns.
The model structure breaks down as follows: decoder stack (4 layers) 193.0M, hidden state projection 21.0M, Markov head 65.5M, and parameters + confidence head 6.4k, totaling 279.5M parameters. This represents only an 8.9% increase over the 3B target model—a low memory cost for significant decoding speedup with no change in output quality.
Benchmark & Competitive Performance
The model was evaluated on six diverse vision tasks under the MMSpec benchmark: general VQA, text VQA, image captioning, chart VQA, complex reasoning, and multi-turn conversation, with a block size of 8. On edge devices with MLX on M5 Max: decoding speedup 2.30x–3.13x and end-to-end latency 1.56x–2.62x. With llama.cpp on M3 Ultra: 1.57x–2.14x decoding and 1.30x–1.77x end-to-end. On H100: 2.04x–2.66x decoding and 1.64x–2.27x end-to-end. The announcement highlights an important constraint under Amdahl's law: speculative decoding accelerates only the decoding phase, not vision encoding or prefill, so on edge devices with limited compute, the end-to-end gain is lower than the decoding-only gain.
Industry Impact & Enterprise Adoption
LFM2.5-VL-DSpark represents a practical opportunity for developers building computer vision applications on edge devices or budget-constrained servers. The additional memory cost of just 8.9% makes deployment on Macs or low-power hardware feasible without sacrificing output quality. The day-one support for llama.cpp, MLX-VLM, and SGLang—along with an OpenAI-compatible endpoint—lowers integration barriers for enterprises already using these frameworks. For the Arabic-speaking developer community, the model's multimodal capabilities and open-weights license enable fine-tuning and deployment for local use cases, though official Arabic benchmarks are not yet available. As speculative decoding gains traction, Liquid AI's approach of pairing a lightweight draft model with a larger target model could become a standard pattern for efficient multimodal inference.
Conclusion
Liquid AI's LFM2.5-VL-DSpark delivers a compelling combination of speed and efficiency: up to 3.13x faster decoding on edge devices and 2.66x on H100, with only 8.9% additional parameters. Its immediate availability on Hugging Face in Safetensors and GGUF formats, along with day-one support for major inference frameworks, positions it as a valuable tool for developers seeking to deploy vision-language models in resource-constrained environments. While Arabic-specific performance remains unverified, the open-weights license and unified representation offer a solid foundation for multilingual adaptation.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
LFM2.5-VL-DSpark is an experimental draft model released by Liquid AI for its LFM2.5-VL-3B vision-language model. It uses speculative decoding to accelerate inference without altering output quality, adding only 280 million parameters (8.9%) to the base model.
It achieves up to 3.13x decoding speedup on edge devices (MLX on M5 Max) and 2.66x on an H100 GPU, with end-to-end latency improvements up to 2.62x and 2.27x respectively. On llama.cpp with M3 Ultra, the speedup ranges from 1.57x to 2.14x.
The model is available on Hugging Face in Safetensors and GGUF formats. It supports inference via llama.cpp (PR#29339), MLX-VLM (PR#2280), and SGLang (PR #40651) from day one, under an open-weights license that permits download, fine-tuning, and deployment without restrictions.
The model is open-weights and free to download from Hugging Face with no licensing fees, meaning the only cost is your own compute hosting. There are no subscription tiers or token prices announced by Liquid AI because distribution is via open weights, not a paid API.
The official announcement does not include a separate Arabic performance evaluation, but the model processes text and images through a unified representation and benefits from decoding acceleration in multilingual tasks. Actual Arabic performance depends on the SFT data, which was not linguistically detailed in the release.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
