Claude Code: connect to a local model when your quota runs out
boxc.net
boxc.net
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 models will move you backwards and just increase the amount of work Claude has to do when your limits reset.
Are there a lot of options how "how far" do you quantize? How much VRAM does it take to get the 92-95% you are speaking of?
So many: https://www.reddit.com/r/LocalLLaMA/comments/1ba55rj/overvie...
> How much VRAM does it take to get the 92-95% you are speaking of?
For inference, it's heavily dependent on the size of the weights (plus context). Quantizing an f32 or f16 model to q4/mxfp4 won't necessarily use 92-95% less VRAM, but it's pretty close for smaller contexts.
Number of params == “variables” in memory
VRAM footprint ~= number of params * size of a param
A 4B model at 8 bits will result in 4GB vram give or take, same as params. At 4 bits ~= 2GB and so on. Kimi is about 512GB at 4 bits.
Now as the other replies say, you should very likely run a quantized version anyway.
https://buildai.substack.com/i/181542049/the-mac-mini-moment
> Thermal throttling: Thunderbolt 5 cables get hot under sustained 15GB/s load. After 10 minutes, bandwidth drops to 12GB/s. After 20 minutes, 10GB/s. Your 5.36 tokens/sec becomes 4.1 tokens/sec. Active cooling on cables helps but you’re fighting physics.
Thermal throttling of network cables is a new thing to me…
I've never heard of this guy before, but I see he's got 5M YouTube subscribers, which I guess is the clout you need to have Apple loan (I assume) you $50K worth of Mac Studios!
I'll be interesting to see how model sizes, capability, and local compute prices evolve.
A bit off topic, but I was in best buy the other day and was shocked to see 65" TVs selling for $300 ... I can remember the first large flat screen TVs (plasma?) selling for 100x that ($30K) when they first came out.
Great demo video though. Nice to see some benchmarks of Exo with this cluster across various models.
No one's running Sonnet/Gemini/GPT-5 locally though.
I'll add on https://unsloth.ai/docs/models/qwen3-coder-next
The full model is supposedly comparable to Sonnet 4.5 But, you can run the 4 bit quant on consumer hardware as long as your RAM + VRAM has room to hold 46GB. 8 bit needs 85.
Kimi K2.5 is a trillion parameter model. You can't run it locally on anything other than extremely well equipped hardware. Even heavily quantized you'd still need 512GB of unified memory, and the quantization would impact the performance.
Also the proprietary models a year ago were not that good for anything beyond basic tasks.
Whether it's a giant corporate model or something you run locally, there is no intelligence there. It's still just a lying engine. It will tell you the string of tokens most likely to come after your prompt based on training data that was stolen and used against the wishes of its original creators.
That might incentivize it to perform slightly better from the get go.
But as a counterpoint: there are whole communities of people in this space who get significant value from models they run locally. I am one of them.
It is probably enough to handle a lot of what people use the big-3 closed models for. Somewhat slower and somewhat dumber, granted, but still extraordinarily capable. It punches way above its weight class for an 80B model.
I need to do more testing before I can agree that it is performing at a Sonnet-equivalent level (it was never claimed to be Opus-class.) But it is pretty cool to get beaten in a programming contest by my own video card. For those who get it, no explanation is necessary; for those who don't, no explanation is possible.
And unlike the hosted models, the ones you run locally will still work just as well several years from now. No ads, no spying, no additional censorship, no additional usage limits or restrictions. You'll get no such guarantee from Google, OpenAI and the other major players.
Hell, if you are willing to go even slower, any GPU + ~80GB of RAM will do it.
Before that I used Qwen3-30B which is good enough for some quick javascript or Python, like 'add a new endpoint /api/foobar which does foobaz'. Also very decent for a quick summary of code.
It is 530Tok/s PP and 50Tok/s TG. If you have it spit out lots of the code that is just copy of the input, then it does 200Tok/s, i.e. 'add a new endpoint /api/foobar which does foobaz and return the whole file'
If you’re worried about others being able to clone your business processes if you share them with a frontier provider then the cost of a Mac Studio to run Kimi is probably a justifiable tax right off.
That said, Claude Code is designed to work with Anthropic's models. Agents have a buttload of custom work going on in the background to massage specific models to do things well.
I was using GLM on ZAI coding plan (jerry rigged Claude Code for $3/month), but finding myself asking Sonnet to rewrite 90% of the code GLM was giving me. At some point I was like "what the hell am I doing" and just switched.
To clarify, the code I was getting before mostly worked, it was just a lot less pleasant to look at and work with. Might be a matter of taste, but I found it had a big impact on my morale and productivity.
I've also been testing OpenClaw. It burned 8M tokens during my half hour of testing, which would have been like $50 with Opus on the API. (Which is why everyone was using it with the sub, until Anthropic apparently banned that.)
I was using GLM on Cerebras instead, so it was only $10 per half hour ;) Tried to get their Coding plan ("unlimited" for $50/mo) but sold out...
(My fallback is I got a whole year of GLM from ZAI for $20 for the year, it's just a bit too slow for interactive use.)
This is a very common sequence of events.
The frontier hosted models are so much better than everything else that it's not worth messing around with anything lesser if doing this professionally. The $20/month plans go a long way if context is managed carefully. For a professional developer or consultant, the $200/month plan is peanuts relative to compensation.
Unless you include it in "frontier", but that has usually been used to refer to "Big 3".
(At todays ram prices upgrading to that for me would pay for a _lot_ of tokens...)
4x rtx 6000 pro is probably the minimum you need to have something reasonable for coding work.
I'd probably rather save the capex, and use the rented service until something much more compelling comes along.
Right now, I think we've probably got 3-5 years of manufacturing woes to work through and another 3-5 years beyond that to get power infrastructure where it needs to be to support it... and even then, I don't think all the resources we can reasonably throw at a combination of mostly nuclear and solar will get there as quickly as it's needed.
That also doesn't consider the bubble itself, or the level of poor to mediocre results altogether even at the frontier level. I mean for certain tasks, it's very close to human efforts in a really diminished timeframe, for others it isn't... and even then, people/review/qa/qc will become the bottleneck for most things in practice.
I've managed to get weeks of work done in a day with AI, but then still have to follow-up for a couple days of iteration on following features... still valuable, but it's mixed. I'm more bullish today than even a few months ago all the same.
Kimi K2.5 is good, but it's still behind the main models like Claude's offerings and GPT-5.2. Yes, I know what the benchmarks say, but the benchmarks for open weight models have been overpromising for a long time and Kimi K2.5 is no exception.
Kimi K2.5 is also not something you can easily run locally without investing $5-10K or more. There are hosted options you can pay for, but like the parent commenter observed: By the time you're pinching pennies on LLM costs, what are you even achieving? I could see how it could make sense for students or people who aren't doing this professionally, but anyone doing this professionally really should skip straight to the best models available.
Unless you're billing hourly and looking for excuses to generate more work I guess?
I still have to occasionally switch to Opus in Opencode planning mode, but not having to rely on Sonnet anymore makes my Claude subscription last much longer.
I'm not looking for a vibe coding "one-shot" full project model. I'm not looking to replace GPT 5.2 or Opus 4.5. But having a local instance running some Ralph loop overnight on a specific aspect for the price of electricity is alluring.
PC or Mac? A PC, ya, no way, not without beefy GPUs with lots of VRAM. A mac? Depends on the CPU, an M3 Ultra with 128GB of unified RAM is going to get closer, at least. You can have decent experiences with a Max CPU + 64GB of unified RAM (well, that's my setup at least).
For VS Code code completion in Continue using a Qwen3-coder 7b model. For CLI work Qwen coder 32b for sidebar. 8 bit quant for both.
I need to take a look at Qwen3-coder-next, it is supposed to have made things much faster with a larger model.
I have claude pro $20/mo and sometimes run out. I just set ANTHROPIC_BASE_URL to a localllm API endpoint that connects to a cheaper Openai model. I can continue with smaller tasks with no problem. This has been done for a long time.
I expect it'll come along but I'm not gonna spend the $$$$ necessary to try to DIY it just yet.
Instead have Claude know when to offload work to local models and what model is best suited for the job. It will shape the prompt for the model. Then have Claude review the results. Massive reduction in costs.
btw, at least on Macbooks you can run good models with just M1 32GB of memory.
Although I'm starting to like LMStudio more, as it has more features that Ollama is missing.
You can then get Claude to create the MCP server to talk to either. Then a CLAUDE.md that tells it to read the models you have downloaded, determine their use and when to offload. Claude will make all that for you as well.
The big powerful models think about tasks, then offload some stuff to a drastically cheaper cloud model or the model running on your hardware.
Which definitely has some questionable implications... but just like with advertising it's not like paying makes the incentives for the people capable of training models to put their thumbs on the scales go away.
Because of how the plugin works in VS code, on my third day of testing with Claude Code, I didn't click the Claude button and was accidentally working with CoPilot for about three hours of torture when I realized I wasn't in Claude Code. Will NEVER make that mistake again... I can only imagine anything I can run at any decent speed lcoally will be closer to the latter. I pretty quickly reach a "I can do this faster/better myself" point... even a few times with Claude/Opus, so my patience isn't always the greatest.
That said, I love how easy it is to build up a scaffold of a boilerplate app for the sole reason to test a single library/function in isolation from a larger application. In 5-10 minutes, I've got enough test harness around what I'm trying to work on/solve that it lets me focus on the problem at hand, while not worrying about doing this on the integrated larger project.
I've still got some thinking and experimenting to do with improving some of my workflows... but I will say that AI Assist has definitely been a multiplier in terms of my own productivity. At this point, there's literally no excuse not to have actual code running experiments when learning something new, connecting to something you haven't used before... etc. in terms of working on a solution to a problem. Assuming you have at least a rudimentary understanding of what you're actually trying to accomplish in the piece you are working on. I still don't have enough trust to use AI to build a larger system, or for that matter to truly just vibe code anything.
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
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)
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
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 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 also highly suggest OpenCode. You'll get the same Claude Code vibe.
If your computer is not beefy enough to run them locally, Synthetic is a bless when it comes to providing these models, their team is responsive, no downtime or any issue for the last 6 months.
Full list of models provided : https://dev.synthetic.new/docs/api/models
Referal link if you're interested in trying it for free, and discount for the first month : https://synthetic.new/?referral=kwjqga9QYoUgpZV
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.,
The rest of your points are why I mentioned AI evals and regressions. I share your sentiment. I've pitched it in the past as "We can’t compare what we can’t measure" and "Can I trust this to run on its own?" and how automation requires intent and understanding your risk profile. None of this is new for anyone who has designed software with sufficient impact in the past, of course.
Since you're interested in combating non-determinism, I wonder if you've reached the same conclusion of reducing the spaces where it can occur and compound making the "LLM" parts as minimal as possible between solid deterministic and well-tested building blocks (e.g. https://alexhans.github.io/posts/series/evals/error-compound... ).
Most Anti-Union arguments I have heard have been about them charging too much in dues, union leadership cozying up to management, and them acting like organized crime doing things like smashing windows of non-union jobs. I have never heard anyone be against unions because they thought they would make it rich on their own.
- I'm pain point driven:
- I can't compare what I can't measure.
- I can't trust to run this "AI" tool to run on its own
Then I treat it as if it was an Integration Test/Test Driven Development exercise of sorts.
- I don't start designing an entire cloud infrastructure.
- I make sure the "agent" is living in the location where the users actually live so that it can be the equivalent of an extra paid set of hands.
- I ask questions or replicate user stories and use deterministic tests wherever I can. Don't just go for LLMaaJ. What's the simplest thing you can think of?
- The important thing is rapid iteration and control. Just like in a unit testing scenario it's not about just writing a 100 tests but the ones that qualitatively allow you to move as fast as possible.
- At this stage where the space is moving so fast and we're learning so much, don't assume or try to over-optimize places that don't hurt and instead think about minimalism, ease of change, parameterization and ease of comparison with other components that form "the black box" and with itself.
- Once you have the benchmarks that you want, you can decide things like pick the cheapest model/agent configuration that does the job within the acceptable timeframe.
Happy to go deeper on these. I have some practical/runnable samples/text I can share on the topic after the weekend. I'll drop a link here when it's ready
Your first two points jive with my intuition that an agents primaries should be a code execution sandbox, mounted files and git.
If you have any practical examples to share I’m sure a ton of people would appreciate it.
It may or may not address the practical examples you need but I'd been to hear your thoughts and maybe it's possible to come up with a more illustrative one.
I didn't go for bubblewrap or similar containers yet because I didn't want to lose a specific type of baseline newcomer yet (Economists who do some coding) but I will be adding to it with whatever most elegant approaches I can find that don't leak too much complexity for things like sandboxing, system testing, integration mocking (reverse proxying), Observing with Openteleletry or otherwise, presenting benchmarks, etc.
1. Switch to extra usage, which can be increased on the Claude usage page: https://claude.ai/settings/usage
2. Logout and Switch to API tokens (using the ANTHROPIC_API_KEY environment variable) instead of a Claude Pro subscription. Credits can be increased on the Anthropic API console page: https://platform.claude.com/settings/keys
3. Add a second 20$/month account if this happens frequently, before considering a Max account.
4. Not a native option: If you have a ChatGPT Plus or Pro account, Codex is surprisingly just as good and comes with a much higher quota.
It’s slower and about 90% as good, so it definitely works as a great back up, but CC with Opus is noticeably better for all of my workloads
This completely depends on the domain, as always. Each of the big 3 have their strengths and weaknesses.
Personally, I’ve used AWS Bedrock as the fallback when my plan runs out, and that seems to work well in my experience. I believe you can now connect to Azure as well.
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?
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.
Some open models have specific training for defined tools (a notable example is OpenAI GPT-OSS and its "built in" tools for browser use and python execution (they are called built in tools, but they are really tool interfaces it is trained to use if made available.) And closed models are also trained to work with generic tools as well as their “built in” tools.
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...
we can do much better with a cheap model on openrouter (glm 4.7, kimi, etc.) than anything that I can run on my lowly 3090 :)
Thanks again for this info & setup guide! I'm excited to play with some local models.
I get that it's not going to work as well as hosted/subscription services like Claude/Gemini/Codex/..., but sometimes those aren't an option
It might make sense to run a small LLM locally for general conversation or very specific tasks, but they're not a serious option for agentic coding.
When I hit the usage limit on the first account, I simply switch to the second and continue working. Since Claude stores progress locally rather than tying it to a specific account, the session picks up right where it left off. That makes it surprisingly seamless to keep momentum without waiting for limits to reset.
Local models are relatively small, it seems wasteful to try and keep them as generalists. Fine tuning on your specific coding should make for better use of their limited parameter count.
I wrote this for the scenario you've run out of quota for the day or week but want a back up plan to keep going to give some options with obvious speed and quality trade-offs. There is also always the option to upgrade if your project and use case needs Opus 4.5.
For example choosing a model that knows rails 8 and react development running on a mac and using docker
Ideally that would make the model small enough to be competitive running locally
Going dumb/cheap just ends up costing more, in the short and long term.
Because mixing company and personal accounts might not be a good idea.
I've been running CC with Qwen3-Coder-30B (FP8) and I find it just as fast, but not nearly as clever.
Will it work? Yes. Will it produce same quality as Sonnet or Opus? No.
Has anyone had a better experience?
Essentially: `ollama launch claude`
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
(Also, I love your podcast!)
But I was really disappointed when I tried to use subagents. In theory I really liked the idea: have Haiku wrangle small specific tasks that are tedious but routine and have Sonnet orchestrate everything. In practice the subagents took so many steps and wrote so much documentation that it became not worth it. Running 2-3 agents blew through the 5 hour quota in 20 minutes of work vs normal work where I might run out of quota 30-45 minutes before it resets. Even after tuning the subagent files to prevent them from writing tests I never asked for and not writing tons of documentation that I didn’t need they still produced way too much content and blew the context window of the main agent repeatedly. If it was a local model I wouldn’t mind experimenting with it more.
This is with my regular $20/month ChatGpT subscription and my $200 a year (company reimbursed) Claude subscription.
Right now OpenAI is giving away fairly generous free credits to get people to try the macOS Codex client. And... it's quite good! Especially for free.
I've cancelled my Anthropic subscription...
Google significantly reduced the free quota and removed pro models from gemini cli some 2-3 moths ago.
Also, Gemini models eat tokens like crazy. Something Codex and Code would do with 2K tokens takes Gemini 100K. Not sure why.
It does eventually finish its quota, but then I just switch to a different Google account (which, amusingly, is what Gemini told me to do).
Happy to consume Google’s free tokens! The free model is a distant third for coding, but it’s fine for leaf node work in a larger project.
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)
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.
LLMs are a hyper-competitive market at the moment, and we have a wealth of options, so if Anthropic is overpricing their API they'll likely be hurting themselves.