Liquid AI

Building a model that fits on a phone or laptop makes pre-training harder. With fewer parameters, quality has to come from better architecture and better data, not just more compute. Our pre-training team, including Aditya Tadimeti and Khoi Nguyen, explains how they design LFMs with the target device in mind from day one.

Today, we are releasing 𝚍𝟷-𝚞𝚕𝚝𝚛𝚊, our most capable decision model yet. Built for hard decision problems, it brings stronger decision quality to complex workflows involving text, images, or both. Across our 112-evaluation suite, d1-ultra leads all eight categories. It supports 44 languages and ranks #1 on the Artificial Analysis Decision Index with a score of 65. A better decision model needs more than accuracy. Its probabilities have to tell teams when to trust an answer. In our tests, d1-ultra had the lowest calibration error of any model we evaluated, making its probabilities more useful for setting automation thresholds and routing uncertain cases for human review. Behind these gains are a larger open-weight backbone, extensive Liquid-specific customizations, and our decision-model training pipeline, with particularly strong improvements on knowledge-related tasks. d1-ultra accepts text, images, or both, along with one question or many. In a single forward pass, it returns a yes/no probability, probabilities across labels, or a score on a scale. These outputs can inform inspection, routing, and review workflows without generating output tokens. d1-ultra is available today through the Liquid API at $0.042 per million input tokens. Text and images are billed at the same input-token rate, with no output-token charges. Each question is billed as its own prompt, including its text and all images. Explore d1-ultra in the d1 Playground, or bring your own inputs and questions to the Liquid API. > Read the blog: https://lnkd.in/gAwiNWUB  > Try the Playground: http://d1.liquid.ai/ > Docs: https://lnkd.in/gHTuKQrp > AA Decision Index: https://lnkd.in/g_sGH6t3  > Start building: https://console.liquid.ai/

Today we release 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱, the largest model in the 𝙻𝙵𝙼𝟸.𝟻 family. It is a 24B-parameter reasoning model with just 2B active parameters per token, designed to work through multi-step tasks inside popular agent harnesses while running efficiently on laptops and PCs. > 24B total parameters, 2B active per token > Context expanded to 131K tokens > A 128K-token vocabulary > 33T pretraining tokens > It is a reasoning model > Very easy to fine-tune and customize On agentic and tool-use tasks, 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱 performs competitively with Qwen3.6-35B-A3B, a model with roughly 50% more total and active parameters. The model is designed for inference efficiency anywhere, with full day-one support for: > llama.cpp and GGUF > MLX for Apple Silicon > vLLM and SGLang 🤝 > ONNX The model's post-training includes reinforcement learning inside agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode. It learns through tool calls and multi-turn tasks, preparing it for agents that need to reason and act. In our demo, 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱 runs locally inside the open-source Hermes Agent desktop application on a single laptop. Model inference stays on-device, without cloud model API calls or model API keys. Start building with 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱. > Blog: https://lnkd.in/gs_FiZza > Docs: https://lnkd.in/gvrc7wxy > Weights on HuggingFace: https://lnkd.in/gYkxdJVU > API: https://console.liquid.ai > Get in touch for insane-mode tuning jobs: https://lnkd.in/gaD9Uw8R

Today we release Open d1: d1-3B and d1-omni-600M, two open-weight models in our d1 decision model family. For teams building applications around text, images, and audio, Open d1 delivers structured decisions in a single forward pass, without generating output tokens. d1-3B handles text and images. Built from LFM2.5-VL-3B, it ranks first among models under 10B on the Decision Index v0.2.1, a benchmark for structured decision-making. It is a strong fit for workflows where decision quality matters, including reranking, agent guardrails, and visual inspection. d1-omni-600M is an experimental 600M-parameter model that accepts text + image or text + audio. It combines an LFM2.5-Encoder-350M backbone with vision and audio encoders and leads our text benchmark comparison in toxicity detection and paraphrase identification. Its compact footprint makes it a fit for voice-command routing, on-device moderation, and intent classification. Both models run across NVIDIA DGX, RTX, and Jetson, with day-one llama.cpp support. For d1-3B, single-question latency measured one request at a time is: > NVIDIA RTX 4090: 8 ms > AMD MI325X: 9.2 ms > NVIDIA Jetson AGX Thor: 16 ms > NVIDIA Jetson AGX Orin 64 GB: 26 ms > Apple M5 Pro: 30 ms > NVIDIA Jetson Orin Nano: 50 ms We built ten live-camera demos powered by d1-3B, from gesture-controlled games to content moderation, with one forward pass per frame. In collaboration with NVIDIA Robotics, we also demonstrate d1-3B navigating an environment in Isaac Sim, with the model served on a Jetson in a hardware-in-the-loop setup. The weights are available on Hugging Face. Download, fine-tune, and deploy locally. > Blog: https://lnkd.in/gaC9jM4S > d1-3B: https://lnkd.in/gqcyY3qr > d1-omni-600M: https://lnkd.in/gYhKqyey > Docs: https://lnkd.in/gHTuKQrp

Liquid 🤝 Arm

Liquid AI is on Fast Company's Next Big Things in Tech list. Our mission is to build efficient general-purpose AI at every scale, bringing advanced intelligence to any processor inside and outside data centers. Glad to see our team’s work recognized. https://lnkd.in/gDiURmQu

𝟭 𝗱𝗮𝘆 𝘁𝗼 𝗹𝗮𝘂𝗻𝗰𝗵. 🚀 At 5:30 the wifi works. Partway through the evening we cut it on purpose, and nothing in the room stops working. ZETIC and Liquid AI are hosting an on-device AI workshop at #SFTechWeek tomorrow. You arrive, get set up, load a model onto your own phone, and then we pull the wifi as a live demo. Everything keeps running, because the models are on the devices. The point is not the party trick. If a model cannot run locally, the product stops working in a basement, a warehouse, a tunnel, or anywhere the signal drops. That is the problem ZETIC works on every day. Both teams will be in the room, so it is a good night to talk to the people actually building this. 𝗦𝗽𝗲𝗮𝗸𝗶𝗻𝗴: Seongjun (SJ) KIM from ZETIC, Viviana Márquez from Liquid AI, and astronomer Brian Castro, who closes the night on telescopes. So the evening goes from models running without the cloud to looking at actual space. Astronaut attire optional but encouraged. Food and drinks on us! Spots are limited and we are close to full. 𝗛𝗼𝘂𝘀𝘁𝗼𝗻, 𝗪𝗲 𝗛𝗮𝘃𝗲 𝗡𝗼 𝗖𝗹𝗼𝘂𝗱 🗓 Tue, Oct 6 · 5:30 to 8:30 PM · San Francisco 💻 Bring a laptop, a cable and a device 🔭 Telescopes after dark RSVP in the comments. #SFTechWeek #OnDeviceAI #ZETIC #LiquidAI

Announcing d1 with vision. 👁️👁️ Our first decision model now supports text, images, or both as inputs. We tested d1 against GPT-6.1 Sol and Claude Opus 5.5 on six real applications, from filtering support tickets to inspecting circuit boards. d1 matches or beats GPT-6.1 Sol on four of them. It costs 19x to 200x less than both models and answers significantly faster on every task. d1 is built for fast, structured decisions: > probabilities for yes/no, choice, or score questions > one forward pass, without generating tokens > text decisions in 200 to 300 ms Vision lets d1 inspect parts directly from camera images: > 85% to 97% accuracy across four VisA inspection tasks > circuit boards, candles, cashews, and chewing gum > understands the task from a short description, without task-specific training d1 can also read game screens and pick the next move: > Tetris: adding the screen raises the number of lines cleared from 70 to 81 > Wordle: solved 12 of 12 games in 3.8 guesses on average, reading the board from screenshots. No text version of the Wordle board is needed. Try the live demos and start building: > d1 Playground: http://d1.liquid.ai/ > Decision models page: https://lnkd.in/gHTuKQrp > Liquid API: https://console.liquid.ai Vercel and OpenRouter support text only for now. Vision is coming to both soon.

LongevityBench, our new aging benchmark with Insilico Medicine, tests models on 17 tasks across clinical records, DNA methylation, transcriptomics, proteomics, and genetics. On several tasks, our compact LFMs outperformed every frontier model we evaluated. What did it take to get there? Katerina Nasto, Liquid AI researcher, and our CTO Mathias Lechner talk through how the benchmark was built, why leakage-proof evaluation matters, and and how specialized models that run locally on sensitive patient data could help researchers move from prediction toward better hypotheses and experiments. We're taking this work to the longevity research community next and will be presenting at Aging Research & Drug Discovery Meeting this week. Reach out if you're attending and want to meet the team! Blog and model: https://lnkd.in/g73eGwJw Paper: https://shorturl.at/qJZLO