- > Reduce your expectations about speed and performance!
Wildly understating this part.
Even the best local models (ones you run on beefy 128GB+ RAM machines) get nowhere close to the sheer intelligence of Claude/Gemini/Codex. At worst these local models will move you backwards and just increase the amount of work Claude has to do when your limits reset.
- The best open models such as Kimi 2.5 are about as smart today as the big proprietary models were one year ago. That's not "nothing" and is plenty good enough for everyday work.
- When your AI is overworked, it gets dumber. It's backwards compatible with humans.
- Useful tip.
From a strategic standpoint of privacy, cost and control, I immediately went for local models, because that allowed to baseline tradeoffs and it also made it easier to understand where vendor lock-in could happen, or not get too narrow in perspective (e.g. llama.cpp/open router depending on local/cloud [1] ).
With the explosion of popularity of CLI tools (claude/continue/codex/kiro/etc) it still makes sense to be able to do the same, even if you can use several strategies to subsidize your cloud costs (being aware of the lack of privacy tradeoffs).
I would absolutely pitch that and evals as one small practice that will have compounding value for any "automation" you want to design in the future, because at some point you'll care about cost, risks, accuracy and regressions.
[1] - https://alexhans.github.io/posts/aider-with-open-router.html
- can you recommend a setup with ollama and a cli tool? Do you know if I need a licence for Claude if I only use my own local LLM?
- What are your needs/constraints (hardware constraints definitely a big one)?
The one I mentioned called continue.dev [1] is easy to try out and see if it meets your needs.
Hitting local models with it should be very easy (it calls APIs at a specific port)
- I've also made decent experiences with continue, at least for autocomplete. The UI wants you to set up an account, but you can just ignore that and configure ollama in the config file
For a full claude code replacement I'd go with opencode instead, but good models for that are something you run in your company's basement, not at home
- we recently added a `launch` command to Ollama, so you can set up tools like Claude Code easily: https://ollama.com/blog/launch
tldr; `ollama launch claude`
glm-4.7-flash is a nice local model for this sort of thing if you have a machine that can run it
- I have been using glm-4.7 a bunch today and it’s actually pretty good.
I set up a bot on 4claw and although it’s kinda slow, it took twenty minutes to load 3 subs and 5 posts from each then comment on interesting ones.
It actually managed to correctly use the api via curl though at one point it got a little stuck as it didn’t escape its json.
I’m going to run it for a few days but very impressed so for for such a small model.
- I think control should be top of the list here. You're talking about building work flows, products and long term practices around something that's inherently non-deterministic.
And the probability that any given model you use today is the same as what you use tomorrow is doubly doubtful:
1. The model itself will change as they try to improve the cost-per-test improves. This will necessarily make your expectations non-deterministic.
2. The "harness" around that model will change as business-cost is tightened and the amount of context around the model is changed to improve the business case which generates the most money.
Then there's the "cataclysmic" lockout cost where you accidently use the wrong tool that gets you locked out of the entire ecosystem and you are black listed, like a gambler in vegas who figures out how to count cards and it works until the house's accountant identifies you as a non-negligible customer cost.
It's akin to anti-union arguments where everyone "buying" into the cloud AI circus thinks they're going to strike gold and completely ignores the fact that very few will and if they really wanted a better world and more control, they'd unionize and limit their illusions of grandeur. It should be an easy argument to make, but we're seeing about 1/3 of the population are extremely susceptible to greed based illusions.,
- Using claude code with custom models
Will it work? Yes. Will it produce same quality as Sonnet or Opus? No.
- Since Llama.cpp/llama-server recently added support for the Anthropic messages API, running Claude Code with several recent open-weight local models is now very easy. The messy part is what llama-server flags to use, including chat template etc. I've collected all of that setup info in my claude-code-tools [1] repo, for Qwen3-Coder-next, Qwen3-30B-A3B, Nemotron-3-Nano, GLM-4.7-Flash etc.
Among these, I had lots of trouble getting GLM-4.7-Flash to work (failed tool calls etc), and even when it works, it's at very low tok/s. On the other hand Qwen3 variants perform very well, speed wise. For local sensitive document work, these are excellent; for serious coding not so much.
One caviat missed in most instructions is that you have to set CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC = 1 in your ~/.claude/settings.json, otherwise CC's telemetry pings cause total network failure because local ports are exhausted.
[1] claude-code-tools local LLM setup: https://github.com/pchalasani/claude-code-tools/blob/main/do...
- God no. "Connect to a 2nd grader when your college intern is too sick to work."
- My experience thus far is that the local models are a) pretty slow and b) prone to making broken tool calls. Because of (a) the iteration loop slows down enough to where I wander off to do other tasks, meaning that (b) is way more problematic because I don't see it for who knows how long.
This is, however, a major improvement from ~6 months ago when even a single token `hi` from an agentic CLI could take >3 minutes to generate a response. I suspect the parallel processing of LMStudio 0.4.x and some better tuning of the initial context payload is responsible.
6 months from now, who knows?
- Open models are trained more generically to work with "Any" tool.
Closed models are specifically tuned with tools, that model provider wants them to work with (for example specific tools under claude code), and hence they perform better.
I think this will always be the case, unless someone tunes open models to work with the tools that their coding agent will use.
- Openrouter can also be used with claude code. https://openrouter.ai/docs/guides/claude-code-integration
- I'm confused, wasn't this already available via env vars? ANTHROPIC_BASE_URL and so on, and yes you may have to write a thin proxy to wrap the calls to fit whatever backend you're using.
I've been running CC with Qwen3-Coder-30B (FP8) and I find it just as fast, but not nearly as clever.
- Or better yet: Connect to some trendy AI (or web3) company's chatbot. It almost always outputs good coding tips
- I gotta say, the local models are catching up quick. Claude is definitely still ahead, but things are moving right along.
- I guess I should be able to use this config to point Claude at the GitHub copilot licensed models (including anthropic models). That’s pretty great. About 2/3 of the way through every day I’m forced to switch from Claude (pro license) to amp free and the different ergonomics are quite jarring. Open source folks get copilot tokens for free so that’s another pro license I don’t have to worry about.
- Opencode has been a thing for a while now
- Or just don’t use Claude Code and use Codex CLI. I have yet to hit a quota with Codex working all day. I hit the Claude limits within an hour or less.
This is with my regular $20/month ChatGpT subscription and my $200 a year (company reimbursed) Claude subscription.
- Or they could just let people use their own harnesses again...
- That wouldn't solve this problem.
And they do? That's what the API is.
The subscription always seemed clearly advertised for client usage, not general API usage, to me. I don't know why people are surprised after hacking the auth out of the client. (note in clients they can control prompting patterns for caching etc, it can be cheaper)
- End users -- people who use harnesses -- have subscriptions so that makes no sense. General API usage is for production.
- "Production" what?
The API is for using the model directly with your own tools. It can be in dev, or experiments, or anything.
Subscriptions are for using the apps Claude + code. That's what it always said when you sign up.
- Production code, of course; deployed software. For when you need to make LLM calls.
- i mean the other obvious answer is to plug in to the other claude code proxies that other model companies have made for you:
https://docs.z.ai/devpack/tool/claude
https://www.cerebras.ai/blog/introducing-cerebras-code
or i guess one of the hosted gpu providers
if you're basically a homelabber and wanted an excuse to run quantized models on your own device go for it but dont lie and mutter under your own tin foil hat that its a realistic replacement
79 points by fugu2 3 days ago | 28 comments