Alternatives
Products that do what Moondream, a small vision language model that runs on 8GB of RAM does
I've been working on training this small vision language model for the last month - excited to release the first prototype today! It is based on SigLIP (image encoder), Phi-1.5 (text model) and trained using the LLaVa-1.5 training dataset. It runs reasonably fast on CPU with ~8GB of RAM in full 32-bit precision. There's plenty of room to speed it up and reduce memory consumption by quantizing the model. I posted a video of it running on my M2 Macbook Air (on CPU not MPS, so performance should be comparable on other hardware) on Twitter to demonstrate inference speed:…
- 1OS
Hi HN, I built a specialized inference engine for running 4-bit Gemma 4 26B-A4B-IT on any M-series Mac using about 2 GB of RAM. It is called TurboFieldfare and is written in Swift and Metal. I have always adored on-device AI. It feels like magic that you can run a powerful NN on your Mac or iPhone. So I wanted to push the limits a bit and run a model whose weights don’t fit in memory. The model’s 4-bit quantized weights occupy roughly 14 GB, which makes running it with conventional inference tools almost impossible on an 8 GB or even 16 GB Mac once the OS, applications, and KV cache are…
Jul 2026 · github.com
- 2AT
A 3.16M-parameter INT4 transformer running entirely in the on-chip memory of a Xilinx Kria KV260. Zero DRAM in the token loop, 59,965 tok/s on the fabric, bit-exact. Chat with it live.
27d ago · mikeayles.com
- 3TA
2020 · thinc.ai
- 4
- 5
The vision plugin for OpenCode that truly understands images. Inspect, read, and reason about any screenshot or picture with deeper understanding than any other plugin — fully local, private, and free. Normally, it takes 300ms to analyse one image on my laptop, which is pretty fast for a local vision model. I use moondream2 as my vision model, you can set your custom model like moondream3.1 if you have a good GPU (for comparison I have currently have an RTX 3050). It works cross-platform. Just follow the README. If you like my work, you leave me a tip as an act for supporting open source!!…
24d ago · github.com
- 6ML
Aug 2026 · github.com
- 7TR
2018 · actcast.io
- 8RG
I wanted to know how fast a 26B mixture-of-experts model could run on a desktop CPU with no GPU. Got ~40 tok/s single-stream (lossless) and ~124 batched. The surprising part was the byte budget: for this model you compress the output head (32% of per-token bytes), not the experts (16%). The writeup has the bandwidth roofline and the dead-ends; the repo has the reproducible recipe. Happy to answer questions. Repo: https://github.com/arun-prasath2005/gemma4-cpu-moe
Jun 2026 · apeg.dev
- 9S1
I wanted to build an inference provider for proprietary AI models, but I did not have a huge GPU farm. I started experimenting with Serverless AI inference, but found out that coldstarts were huge. I went deep into the research and put together an engine that loads large models from SSD to VRAM up to ten times faster than alternatives. It works with vLLM, and transformers, and more coming soon. With this project you can hot-swap entire large models (32B) on demand. Its great for: Serverless AI Inference Robotics On Prem deployments Local Agents And Its open source. Let me know if anyone…
Nov 2025 · github.com
- 10IR
Jun 2026 · github.com
- 11W1
Hey HackerNews, I built this project over the last few weeks as a palette cleanser from a failed game launch. I wanted to learn a bit about AI/Neural-Networks and naively thought I could build a tiny maze-solving AI in a weekend with a 100% solve rate. Well - I couldn't, but I got pretty close. 14 Bytes total model size, and a 96.5% solve rate on unseen mazes. Trained across 46 phases experimenting with different ideas to improve the model (better performance, smaller size). Its quite fun to watch the model attempt to solve the maze, when they fail its usually due to getting stuck in a…
Jul 2026 · con-dog.github.io
- 12PA
Hello Hacker News! I am Bertrand from Pruna AI. With my associates, John, Rayan, and Stephan, we are fellow researchers in AI efficiency and reliability coming from TUM. We are building an optimization engine that combines compression methods (e.g. quantization, pruning, compilation, batching…) in the aim of saving compute power when running AI models. This optimization engine take one base model as input and returns a compressed model as output. It aims to help for two things: - Make various AI models faster and/or smaller for various hardware (because they can require significant…
2024
- 13MM
Hi HN! We (Thomas and Stéphan, hello!) recently released Model2Vec, a Python library for distilling any sentence transformer into a small set of static embeddings. This makes inference with such a model up to 500x faster, and reduces model size by a factor of 15 (7.5M params or 15/30MB on disk, depending on whether you use float16 or float32). This allows you to embed 50-100k documents per second on a cpu on a macbook. This reduction of course comes at a cost: distilled models are worse than their parent models. Even so, they are actually a lot better than large sets of conventional…
2024 · github.com
- 14AN
Kimi K3 has 2.78 trillion parameters and ships as 1.42 TB of weights. It clearly does not fit in the memory of a laptop. But K3 is a Mixture-of-Experts model. For each token, only a small fraction of its 896 experts per layer is activated. That changes the problem: the entire model does not need to be resident in RAM, as long as the weights required by each token can be reached quickly enough. We built WASTE — the Weight-Aware Streaming Tensor Engine — to explore that idea. WASTE keeps the dense, repeatedly used part of the model resident in memory, stores the routed experts in an…
Jul 2026
- 15IC
I spent the past week implementing a 1 Layer Neural Net and training it on MNIST within the visual scripting language provided by scratch.mit.edu. It was tedious, but ultimately not too difficult. The code runs incredibly slowly, so much so that 64 samples of MNIST takes 5+ hours to train on my machine. There were a lot of little mini challenges that were fun to overcome (implementing softmax was very tricky). If you're interested, I encourage you to try and improve on it! More details in the linked blog post.
2024 · bell-boy.github.io
- 16CT
I had been looking to try <500M parameter language models but you wouldn't find an API to try them anywhere, so I built this cloudflare hosted static website that hosts weights and built an inference runtime for these models that uses WebGPU and runs inference from your browser. These are only so useful in a multi-turn conversation but it's still interesting to see what you can pack in a <250mb model. I tried using ONNX versions earlier, but there were too many quirks of using them with language models and the TPS wasn't too impressive. Inspired by svenflow/webgpu-gemma, I put my codex…
May 2026 · chonklm.com
- 17SV
2017 · github.com
- 18OA
A friend and I are launching an alpha first thing in 2011. We're trying to gain some momentum and whatnot, so we're opening early registration as of tonight. We're planning to develop a Bayesian network to derive suggestions. If you're interested in seeing our progress, check it out. Edit: Any and all suggestions, criticism, advice, etc is highly appreciated!
2010 · osmoar.com
- 195L
We've built InferX, a specialized runtime environment that fundamentally changes how LLMs are served. The core problem we solve is the latency bottleneck in AI inference, especially with large models. Current systems waste resources or suffer from painfully slow cold starts. InferX's AI-native architecture, with its "snapshot" technology, enables: * *Sub-2s cold starts:* Spin up models instantly. * *High density:* Serve more LLMs on the same GPUs. * *Optimal efficiency:* Maximize GPU utilization. This isn't just another API; it's a new execution layer designed from the ground up for the…
2025 · github.com
- 20IM
Hey HN! Thank you for all the support and feedback on my original submission 2 months ago. I've been improving the backend using a MCTS/AlphaZero approach and it's currently producing much better results. My long term goal is to allow users to manage multiple projects, deployed autonomously, both from scratch and by making continual updates all prompted with natural language. The cost of each project has been lowered to $9 as performance with smaller models has improved (I migrated from Claude-3-Opus to gemini-1.5-flash). Thanks for checking it out!
2024 · saas-quick.com
- 21BO
Read the full blogpost at https://rach.codes/blog/Introducing-Bhumi (click on reader to see the technical breakdown!) AI inference should be fast, but in practice it’s painfully slow. Inference bottlenecks slow down LLM-powered chatbots and AI workflows everywhere. I built Bhumi to fix that. Bhumi is a Python library designed for developers, yet its performance-critical core is implemented in Rust (via PyO3) for near-native speed. This hybrid approach delivers up to 2.5x faster response times across providers like OpenAI, Anthropic, and Gemini—without changing the…
2025 · bhumi.trilok.ai
- 22IB
I had 14,000 photos sitting on a drive and wanted an excuse to play with local vision models and Elixir/Phoenix. I originally tried to get LLaVA to tell me if a photo was 'good' or matched my style, but quickly learned that LLMs have terrible taste. I ended up demoting the LLM to just extract metadata, and built a custom CLIP/Ridge Regression pipeline to actually learn my preferences based on how I rate things. The stack is Phoenix/Oban on the orchestrator side, and Python/FastAPI/Instructor for the AI workers. Happy to answer any questions about the architecture,…
Apr 2026 · qwelian.com
- 23TN
Hi guys, I’m excited to share an update on ReproModel, an open-source toolbox designed to streamline the testing and reproduction of machine learning models. I, like many of you, have really struggled with benchmarking and comparing models, from missing code, to opaque experiment parameters slowing the process. I decided to take matters into my own hands, and created a mini-toolbox in my free time to streamline the process. The goal is to reduce the time and effort spent on replicating experiments, enabling researchers to focus on innovation rather than setup. Knowing this task is not an…
2024 · github.com
- 24MS
I’m not a software engineer or a genius — I just had a weird idea: What if memory wasn’t just stored as text or embeddings, but as symbolic, byte-level thoughts that could be passed between AIs? That idea became MemoryCore Lite: Encodes thoughts into lightweight bytecode Shares them across nodes via peer-to-peer sync Fully decentralized, no GPU needed Designed to evolve into its own AI knowledge mesh I just open-sourced the basic version here: github.com/ProToxicNinja/MemoryCore-Lite-Symbolic-Memory-Engine-for-AI It’s simple — but everything works. You can build better tokenizers,…
2025 · github.com
Ranked by how close each launch is in meaning, then by votes. Refine with a description →