Comment by jedbrooke
13 hours ago
I’d been thinking about if something like this would be possible for https://chatjimmy.ai/ . The underlying model is only llama 3 8B but I’m curious what coding harnesses would be like at 17k tok/s
13 hours ago
I’d been thinking about if something like this would be possible for https://chatjimmy.ai/ . The underlying model is only llama 3 8B but I’m curious what coding harnesses would be like at 17k tok/s
If you're on macOS you can try the built in LLM which I think is similar in size. There's a project called Apfel that wraps it in a CLI. Also Chrome ships with a web API called Prompt API that gives you offline access to Gemini Nano which can do both text and images at the input. Also tiny. I've integrated these into my workflows where a tiny but non zero amount of reasoning is needed in between the otherwise fully deterministic steps.
looks like the macOS one is Tahoe only. I’ve been putting of upgrading to tahoe but this might be enough to tempt me
What kind of reasoning makes this worthwhile?
I have a personal, fully offline and local version of Windows Recall basically, but good, made using macOS built-in OCR and LLM. The reasoning requirements are tiny (just interpret the screen based on the OCR, do rolling de-duplication and summarization), but they are non-zero. The tool is valuable to me and it being dep-free and fully offline and local just gives me a good feeling.
2 replies →
I actually tried building a harness around their constraints, just to find out if it was possible, but the combination of small context window, no tool calls and just small model, made me understand, that it’s not going to work.
If you find a way to do it, I’d love to hear it!
I added it in my oh-my-pi configuration before (it's OpenAI compatible), but Llama 3 8B is just absolutely unusable for anything coding related. It is very fast and the latency is very good however.
I tried the site and can't find any information about what it is. What is it?
They make custom chips with a model's weights and parameters "hard-coded" which allows for much, much faster inference.
Codex offers a -spark model that runs on Cerebras. Not quite 17k tok/s, but _very_ fast nonetheless. Worth a look.