LiquidAI d1-3B is a 3-billion-parameter multimodal decision model packaged for image-text-to-text tasks and optimized for classification, calibration, and conversational decision-making. Built on LiquidAI’s LFM2.5-VL-3B base, it supports 16 languages and is tagged for edge deployment, system-one workflows, and custom-code integrations. Model artifacts are provided in safetensors (BF16 parameter count ~3.12B; total files around 6.25 GB), licensed under lfm1.0, region-marked US, and marked compatible with inference endpoints and quantizations (hasQuantizations true, isQuantized false). Metadata highlights links to the LFM2 technical report (arXiv:2511.23404), 23 likes and 15 downloads, and pipeline classification labels emphasizing decision, calibration, and multimodal conversational use.
Usage instructions target practical deployment: examples show how to run an image-text-to-text pipeline with Transformers (pipeline("image-text-to-text"), AutoProcessor and AutoModelForMultimodalLM) and how to serve via vLLM with an OpenAI-compatible API. Example inputs illustrate chat-style messages combining image URL objects and text prompts and recommend device_map/auto and processor.apply_chat_template before model.generate. The package is positioned for researchers and engineers who need a multilingual, edge-capable multimodal decision model with ready integration into common tooling (Transformers, vLLM) and endpoint workflows.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
An AI developer has launched open-source, Adobe-like creative software suites using AI reverse engineering, aiming for full feature parity. The projects, built with Anthropic’s Claude Opus, are in early alpha but plan to reach near-complete functionality within months. (arstechnica.com)
A new open vision-language model has been developed for various medical applications, enabling improved interpretation of medical data. The model aims to enhance diagnostics and treatment through advanced AI understanding of medical images and language. (nature.com)
OpenAI released 372 mathematical results generated by an internal model, many of which are verified by formal proofs. These proofs are unfamiliar and uncanny, reflecting human knowledge but in a form humans may not fully understand. (avowed.substack.com)
Google AI Edge Foresight enables offline, private transcription of meeting content. It aims to enhance data privacy and accessibility for users. (developers.google.com)
Meta and Microsoft are reducing employee use of Anthropic’s Claude AI, shifting focus to their own AI tools. Microsoft is cutting internal AI spending, while Meta is replacing Claude with its proprietary models for internal use. (rswebsols.com)
Qwen 3.8-27B, a free local language model, replaces paid cloud AI services like ChatGPT and Claude for a user’s daily tasks. It provides similar performance at no cost, removing the need for subscriptions. (xda-developers.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.