Show HN: docker/model-runner – an open-source tool for local LLMs
Show HN (score: 16)Description
We're the maintainers of docker/model-runner and wanted to share some major updates we're excited about.
Link: https://github.com/docker/model-runner
We are rebooting the community:
https://www.docker.com/blog/rebooting-model-runner-community...
At its core, model-runner is a simple, backend-agnostic tool for downloading and running local large language models. Think of it as a consistent interface to interact with different model backends. One of our main backends is llama.cpp, and we make it a point to contribute any improvements we make back upstream to their project. It also allows people to transport models via OCI registries like Docker Hub. Docker Hub hosts our curated local AI model collection, packaged as OCI Artifacts and ready to run. You can easily download, share, and upload models on Docker Hub, making it a central hub for both containerized applications and the next wave of generative AI.
We've been working hard on a few things recently:
- Vulkan and AMD Support: We've just merged support for Vulkan, which opens up local inference to a much wider range of GPUs, especially from AMD.
- Contributor Experience: We refactored the project into a monorepo. The main goal was to make the architecture clearer and dramatically lower the barrier for new contributors to get involved and understand the codebase.
- It's Fully Open Source: We know that a project from Docker might raise questions about its openness. To be clear, this is a 100% open-source, Apache 2.0 licensed project. We want to build a community around it and welcome all contributions, from documentation fixes to new model backends.
- DGX Spark day-0 support, we've got it!
Our goal is to grow the community. We'll be here all day to answer any questions you have. We'd love for you to check it out, give us a star if you like it, and let us know what you think.
Thanks!
More from Show
Show HN: Zenode – an AI-powered electronic component search engine
Show HN: Zenode – an AI-powered electronic component search engine TL;DR - My cofounder Collin and I built an AI version of Digi-Key to help PCB designers find and use parts, except with a way bigger catalog, modern refinement tools, and an AI that can actually read the damn datasheets for you.<p>*The problem*<p>Modern circuit board design is filled with absurdly tedious tasks, where one small mistake can brick a project and cost thousands. The worst (in our opinion) is reading datasheets, which eats up to 25% of the first part of any project: 1. First, you slog through catalogs to find viable parts, using search tools that are still stuck in the dark ages. There are ~80M unique components in today’s supply chain, yet the tools we have to look through them are just digitized versions of the same paper catalogs our grandparents got in the mail.<p>2. During the design, you spend a ton of time flipping between different 10-100-page PDFs for every component in every subcircuit, hoping like hell you don’t miss some tiny spec in a footnote somewhere that kills your design.<p>3. And god help you when the requirements inevitably change and now you have to figure out what subsystems are affected!<p>*What we built*<p>Zenode is an AI-powered electronics search engine that actually helps engineers find and understand components. Our core features: 1. Largest and Deepest Part Catalog → We have merged dozens of existing part catalogs and documents from major distributors and manufacturers<p>2. Discovery Search → natural language queries to quickly find categories, set filters, and rank results<p>3. Modern Parametric Filters → rebuilt from scratch to move off the string values pervasive in industry and build numeric ranges that actually work.<p>4. Interactive Documents → AI constrained to a single part’s datasheet/manuals. Ask a question, get the answer with a highlighted source for quick reference.<p>5. Deep Dive → search across dozens of parts simultaneously (“what’s the lowest-power accelerometer available?”) instead of slogging one by one.<p>*What we learned*<p>1. By far the hardest part of the last 2 years has been wrangling 3 TB of messy, inconsistent data into something usable. We had to teach the AI how to handle hand-drawn figures, normalize different unit variables and names that mean the same thing, and navigate conflicting information present between different datasheet versions of the same part. It’s been a nightmare<p>2. We originally built custom PDF parsers and AI extractors, which were best in class for ~3 months until generalized AI passed them. So we stopped reinventing wheels and doubled down on data quality instead.<p>3. The killer feature wasn’t the AI searching a single part, but what we heard repeatedly from users is that they want the AI to read across multiple parts, hence why we’ve launched deep dive!<p>*Where it’s strong*<p>- Speed: rips through a 1,000-page microcontroller datasheet in seconds.<p>- Breadth: 40M+ part sources unified into one catalog, and more than just datasheets, application notes, errata, etc.<p>- Comparisons: Deep Dive lets you ask across multiple parts, not just one at a time.<p>*Where it’s not*<p>- Pricing/availability: currently outdated (for now we expect folks to check existing aggregators like Octopart).<p>- Accuracy: good enough to match my mediocre skills; not yet at Collin's level, but we're starting tuning and this will improve rapidly!<p>*Try it*<p>It’s live today (zenode.ai). Sign up for a free account and If you put “Hacker News” in during signup in the “where did you hear about us” field, we’ll give you 1,000 bonus credits (once we finish building that, so sometime this week ).<p>*Feedback we’d love*<p>1. Should Deep Dive results auto-become filters you can refine further?<p>2. Do you want the ability to mark preferred parts / exclude others?<p>3. Is “Deep Dive on a BOM” (alt discovery + manufacturability checks on a list of known components from different categories) the killer feature?
Show HN: Nallely – A Python signals/MIDI processing system inspired by Smalltalk
Show HN: Nallely – A Python signals/MIDI processing system inspired by Smalltalk Nallely is about experimenting with signals: routing, patching, or writing small neurons that process signals and eventually sink in MIDI devices or any application connected to Nallely.<p>I try to get inspired by the "Systems as a Living Things" philosophy and aim, step by step, to create an auto-adaptive, resilient, distributed system. Currently, neurons live in their own thread in a session (world), and send signals (messages) to each other through patches (channels). You can also connect to a network-bus neuron to register your own neurons written in any other technology and have them interact with the existing neurons inside the world. Nallely offers an API to easily code your own reactive neurons, and provides a mobile-friendly GUI for patching everything visually.<p>As anyone posting something based on Python, I can already hear: "no, Python's bad, think about the performances, think about the children".<p>We all know about Python performances (we've all seen the animation with the moving balls and stuff), but the focus here is on dynamic and emergent behaviors, extensibility, and run time adaptability over extreme performance. Even though Nallely is written in pure Python, it runs on a Raspberry Pi 5 (ok, a powerful one), consuming less than 10% CPU on a normal usage and around 40MB of memory.<p>And, as someone mentioning Smalltalk, I can already hear: "Why didn't you write it in Smalltalk"? (replace Smalltalk by your prefered dialect)<p>I like Smalltalk, but I also like Python. Nailed it, perfect justification. Jokes aside, IMO Smalltalk is "Systems as Living Things" pushed at its extreme for designing a language, and I admire that. With Nallely, I want to explore the same philosophy: independent musical/signal-processing neurons, without relying on Smalltalk, while benefiting from Python's deployment and ecosystem advantages (compared to Smalltalk).
Show HN: Robot MCP Server – Connect Any Language Model and ROS Robots Using MCP
Show HN: Robot MCP Server – Connect Any Language Model and ROS Robots Using MCP We’ve open-sourced the Robot MCP Server, a tool that lets large language models (LLMs) talk directly to robots running ROS1 or ROS2.<p>What it does - Connects any LLM to existing ROS robots via the Model Context Protocol (MCP) - Natural language → ROS topics, services, and actions (And the ability to read any of them back) - Works without changing robot source code<p>Why it matters - Makes robots accessible from natural language interfaces - Opens the door to rapid prototyping of AI-robot applications - We are trying to create a common interface for safe AI ↔ robot communication<p>This is too big to develop alone — we’d love feedback, contributors, and partners from both the robotics and AI communities.
No other tools from this source yet.