Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models. A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality. Across six vision-language task categories at batch size 1 and temperature 0: > MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end > llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end > SGLang on H100: up to 2.66x faster decoding and 2.27x end to end All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data. 🧵

Started at an @AnthropicAI Hackathon in February. Today, Matrix OS (@joinmatrixos) is live, giving a computer for your AI agents so they don't die when you close your laptop. But for me, the story started much earlier. When I was six, I could barely read, but I would spend hours clicking through Windows XP just trying to understand what a computer could do. I think I’ve been obsessed with that question ever since. Over the past 3.5 years, I’ve built intensively with AI. I kept seeing the same pattern: the models were getting incredibly capable, but the computer around them was still designed for a human sitting in front of a screen, clicking around and typing. I joined forces with @NimaNaderi2004, we kept building, and today we’re opening signups, shipping Matrix and our desktop app, and announcing our first angel round, backed by @alexwikstrom, @bakkermichiel, @salar, @Founder_Ash, @mattiasmiksche, @fbjork, @sidk_, and @engineerrprompt among others. Since we are backed by the best, we brought @xyhstruggler , @Elibethmarch and @valeriewistrom to help build and achieve one mission. Our mission is to give everyone a Personal Cloud Computer 💚 PS: We’re also hiring engineers in Stockholm and around the world. If rethinking what a computer can be; sounds like your kind of problem, come build with us!