Live data from Hacker News

A guide to local coding models

aiforswes.com

231–240 of 363 posts

Re: A guide to local coding models

#231
post #110

Earlier quoted context omitted.

As a third option, I've found I can do a few hours a day on the $20/mo Google plan. I don't think Gemini is quite as good as Claude for my uses, but it's good enough and you get a lot of tokens for your $20. Make sure to enable the Gemini 3 preview in gemini-cli though (not enabled by default).

Huge caveat: For the $20/mo subscription Google hasn't made clear if they train on your data. Anthropic and OAI on the other hand either clearly state they don't train on paid usage or offer very straightforward opt-outs. https://geminicli.com/docs/faq/ > What is the privacy policy for using Gemini Code Assist or Gemini CLI if I’ve subscribed to Google AI Pro or Ultra? > To learn more about your privacy policy and te…

That's the main reason, why I hope Google does not win this AI war.

Re: A guide to local coding models

#232
post #217
post #203

I am still hoping, but for the moment… I have been trying every 30-80B model that came out in the last several months, with crush and opencode, and it's just useless. They do produce some output, but it's nowhere near the level that claude code gets me out of the box. It's not even the same league. With LLMs, I feel like price isn't the main factor: my time is valuable, and a tool that doesn't improve the way I work…

Claude Code is a lot about prompting and orchestration of the conversation. The LLM is just a tool in these agentic frameworks. Whats truly ingenious is how context is engineered/managed, how is the code-RAG approached, and them LLM memory that is used. So my guess would be - we need open conversation or something along the line of "useful linguistic-AI approaches for combing and grooming code"

Agreed. I've been trying to use opencode and crush, and none of them do anything useful for me. In contrast, claude code "just works" and does genuinely useful work. And it's not just because of the specific LLM used, it's the overall engineering of the tool, the prompt behind the scenes, etc.

But the bottom line is that I still can't find a way to use either local LLMs and/or opencode and crush for coding.

Re: A guide to local coding models

#234

Earlier quoted context omitted.

I've been a software developer for 25 years, and 30ish years in the industry, and have been programming my whole life. I worked at Google for 10 of those years. I work in C++ and Rust. I know how to write code. I don't pay $100 to "vibe code" and "learn to program" or "avoid learning to program." I pay $100 so I can get my personal (open source) projects done faster and more completely without having to hire people w…

Why would you ever hire someone to help with a personal open source project?

because we want to support open source? Even if you're independence maximalist, you still pay other people in your life to do things for you at some point. If you've got the money and the desire but not the time, why does that not seem reasonable to you?

Re: A guide to local coding models

#235

Earlier quoted context omitted.

This older HN thread shows R1 running on a ~$2k box using ~512 GB of system RAM, no GPU, at ~3.5-4.25 TPS: https://news.ycombinator.com/item?id=42897205 If you scale that setup and add a couple of used RTX 3090s with heavy memory offloading, you can technically run something in the K2 class.

Stop recommending 3090s they are all but obsolete now. Not having native bf16 is a showstopper.

Hard disagree. The difference in performance is not something you'll notice if you actually use these cards. In AI benchmarks, the RTX 3090 beats the RTX 4080 SUPER, despite the latter having native BF16 support. 736GiB/s (4080) memory bandwidth vs 936 GiB/s (3090) plays a major role. Additionally, the 3090 is not only the last NVIDIA consumer card to support SLI.

It's also unbeatable in price to performance as the next best 24GiB card would be the 4090 which, even used, is almost tripple the price these days while only offering about 25%-30% more performance in real-world AI workloads.

You can basically get an SLI-linked dual 3090 setup for less money than a single used 4090 and get about the same or even more performance and double the available VRAM.

Re: A guide to local coding models

#236
post #110

Earlier quoted context omitted.

As a third option, I've found I can do a few hours a day on the $20/mo Google plan. I don't think Gemini is quite as good as Claude for my uses, but it's good enough and you get a lot of tokens for your $20. Make sure to enable the Gemini 3 preview in gemini-cli though (not enabled by default).

Huge caveat: For the $20/mo subscription Google hasn't made clear if they train on your data. Anthropic and OAI on the other hand either clearly state they don't train on paid usage or offer very straightforward opt-outs. https://geminicli.com/docs/faq/ > What is the privacy policy for using Gemini Code Assist or Gemini CLI if I’ve subscribed to Google AI Pro or Ultra? > To learn more about your privacy policy and te…

I have no idea at all whether the GCP "Service Specific Terms" [1] apply to Gemini CLI, but they do apply to Gemini used via Github Copilot [2] (the $10/mo plan is good value for money and definitely doesn't use your data for training), and states:

  Service Terms
  17. Training Restriction. Google will not use Customer Data to train or fine-tune any AI/ML models without Customer's prior permission or instruction.
[1] https://cloud.google.com/terms/service-terms

[2] https://docs.github.com/en/copilot/reference/ai-models/model...

Re: A guide to local coding models

#237

Earlier quoted context omitted.

I use local models + openrouter free ones. My monthly spend on ai models is I'm not cheap, just ahead of the curve. With the collapse in inference cost, everything will be this eventually I'll basically do $ man tool | or even $ cat source | Things I used to do intensively I now do lazily. I've even made a IEITYuan/Yuan-embedding-2.0-en database of my manpages with chroma and then I can just ask my local documentatio…

> I'll basically do $ man tool | or even $ cat source | Could you please elaborate on this? Do I get this right that you can set up your your command line so that you can pipe something to a command that sends this something together with a question to an LLM? Or did you just mean that metaphorically? Sorry if this is a stupid question.

Yes, I use simonw's `llm` for that: https://github.com/simonw/llm

Example:

    $ man tar | llm "how do I extract test.txt from a tar.gz"

Re: A guide to local coding models

#238
My experience: even for the run of the mill stuff, local models are often insufficient, and where they would be sufficient, there is a lack of viable software.

For example, simple tasks CAN be handled by Devstral 24B or Qwen3 30B A3B, but often they fail at tool use (especially quantized versions) and you often find yourself wanting something bigger, where the speed falls a bunch. Even something like zAI GLM 4.6 (through Cerebras, as an example of a bigger cloud model) is not good enough for doing certain kinds of refactoring or writing certain kinds of scripts.

So either you use local smaller models that are hit or miss, or you need a LOT of expensive hardware locally, or you just pay for Claude Code, or OpenAI Codex, or Google Gemini, or something like that. Even Cerebras Code that gives me a lot of tokens per day isn't enough for all tasks, so you most likely will need a mix - but running stuff locally can sometimes decrease the costs.

For autocomplete, the one thing where local models would be a nearly perfect fit, there just isn't good software: Continue.dev autocomplete sucks and is buggy (Ollama), there don't seem to be good enough VSC plugins to replace Copilot (e.g. with those smart edits, when you change one thing in a file but have similar changes needed like 10, 25 and 50 lines down) and many aren't even trying - KiloCode had some vendor locked garbage with no Ollama support, Cline and RooCode aren't even trying to support autocomplete.

And not every model out there (like Qwen3) supports FIM properly, so for a bit I had to use Qwen2.5 Coder, meh. Then when you have some plugins coming out, they're all pretty new and you also don't know what supply chain risks you're dealing with. It's the one use case where they could be good, but... they just aren't.

For all of the billions going into AI, someone should have paid a team of devs to create something that is both open (any provider) and doesn't fucking suck. Ollama is cool for the ease of use. Cline/RooCode/KiloCode are cool for chat and agentic development. OpenCode is a bit hit or miss in my experience (copied lines getting pasted individually), but I appreciate the thought. The rest is lacking.

Re: A guide to local coding models

#239

Earlier quoted context omitted.

The limits for the $20/month plan can be reached in 10-20 minutes when having it explore large codebases with directed. It’s also easy to blow right through the quota if you’re not managing content well (waiting until it fills up and then auto-compacting, or even using /compact frequently instead of /clear or the equivalent in different tools). For most of my work I only need the LLM to perform a structured search of…

Not for Codex. Not even for Gemini/Antigravity! I am truly shocked by how much mileage I can get out of them. I recently bought the $200/mo OpenAI subscription but could barely use 10% of it. Now for over a month, I use codex for at least 2 hrs every day and have yet to reach the quota. With Gemini/Antigravity, there’s the added benefit of switching to Claude Code Opus 4.5 once you hit your Gemini quota, and Google i…

Can confirm this is the way right now
Post reply on HN