I've run several local models that get this right. Qwen 3.5 122B-A10B gets this right, as does Gemma 4 31B. These are local models I'm running on my laptop GPU (Strix Halo, 128 GiB of unified RAM).
And I've been using this commonly as a test when changing various parameters, so I've run it several times, these models get it consistently right. Amazing that Opus 4.7 whiffs it, these models are a couple of orders of magnitude smaller, at least if the rumors of the size of Opus are true.
I'm running an 8 bit quant right now, mostly for speed as memory bandwidth is the limiting factor and 8 bit quants generally lose very little compared to the full res, but also to save RAM.
I'm still working on tweaking the settings; I'm hitting OOM fairly often right now, it turns out that the sliding window attention context is huge and llama.cpp wants to keep lots of context snapshots.
I had a whole bunch of trouble getting Gemma 4 working properly. Mostly because there aren't many people running it yet, so there aren't many docs on how to set it up correctly.
It is a fantastic model when it works, though! Good luck :)
You may be right that they were political in that sense.
But also, they probably should have happened were he not a politician. He's been committing fraud and other white collar crimes for quite a while. Unfortunately, we go far too easy on white collar crime in this country. And he's a master of plausible deniability, where he effectively asks other people to commit crimes on his behalf, but in a plausibly deniable way with no written trail.
Unfortunately, very few LLM benchmarks do this. LLMs get such high scores on many benchmarks because there's no difference between answering "I don't know" as giving a made up answer, and made up answers can improve the score some of the time, so by chasing higher benchmark numbers on these kinds of benchmarks, the labs are prioritizing guessing over accuracy.
The Artificial Analysis Omniscience benchmark does penalize guessing, so it actually helps you determine which LLMs are likely to just guess rather than telling you they don't know. Only a very few of the frontier models actually score higher than 0 on this, where 0 means that it's equally likely to return a correct answer as it is to return a hallucination on factual questions.
I guess one of the things I don't understand: how you expect a stochastic model, sold as a proprietary SaaS, with a proprietary (though briefly leaked) client, is supposed to be predictable in its behavior.
It seems like people are expecting LLM based coding to work in a predictable and controllable way. And, well, no, that's not how it works, and especially so when you're using a proprietary SaaS model where you can't control the exact model used, the inference setup its running on, the harness, the system prompts, etc. It's all just vibes, you're vibe coding and expecting consistency.
Now, if you were running a local weights model on your own inference setup, with an open source harness, you'd at least have some more control of the setup. Of course, it's still a stochastic model, trained on who knows what data scraped from the internet and generated from previous versions of the model; there will always be some non-determinism. But if you're running it yourself, you at least have some control and can potentially bisect configuration changes to find what caused particular behavior regressions.
The problem is degradation. It was working much better before. There are many people (some example of a well know person[0]), including my circle of friends and me who were working on projects around the Opus 4.6 rollout time and suddenly our workflows started to degrade like crazy. If I did not have many quality gates between an LLM session and production I would have faced certain data loss and production outages just like some famous company did. The fun part is that the same workflow that was reliably going through the quality gates before suddenly failed with something trivial. I cannot pinpoint what exactly Claude changed but the degradation is there for sure. We are currently evaling alternatives to have an escape hatch (Kimi, Chatgpt, Qwen are so far the best candidates and Nemotron). The only issue with alternatives was (before the Claude leak) how well the agentic coding tool integrates with the model and the tool use, and there are several improvements happening already, like [1]. I am hoping the gap narrows and we can move off permanently. No more hoops, you are right, I should not have attempted to delete the production database moments.
Curious as to how many people are using 4.6, perhaps you’re on a subscription? I use the api and 4.6 (also goes for Sonnet) is unusable since launch because it eats through tokens like it’s actually made that way (to make more money/hit limits faster). I guess it makes sense from a financial perspective but once 4.5 goes away I will have to find another provider if they continue like this :/
If you get consistently nowhere near 50% then surely you know you're not throwing a fair coin? What would complaining to the coin provider achieve? Switch coins.
Yes, you can use it for local coding. Most harnesses can be pointed at a local endpoint which provides an OpenAI compatible API, though I've had some trouble using recent versions of Codex with llama.cpp due to an API incompatibility (Codex uses the newer "responses" API, but in a way that llama.cpp hasn't fully supported).
I personally prefer Pi as I like the fact that it's minimalist and extensible. But some people just use Claude Code, some OpenCode, there are a ton of options out there and most of them can be used with local models.
Right now, open models that run on hardware that costs under $5000 can get up to around the performance of Sonnet 3.7. Maybe a bit better on certain tasks if you fine tune them for that specific task or distill some reasoning ability from Opus, but if you look at a broad range of benchmarks, that's about where they land in performance.
You can get open models that are competitive with Sonnet 4.6 on benchmarks (though some people say that they focus a bit too heavily on benchmarks, so maybe slightly weaker on real-world tasks than the benchmarks indicate), but you need >500 GiB of VRAM to run even pretty aggressive quantizations (4 bits or less), and to run them at any reasonable speed they need to be on multi-GPU setups rather than the now discontinued Mac Studio 512 GiB.
The big advantage is that you have full control, and you're not paying a $200/month subscription and still being throttled on tokens, you are guaranteed that your data is not being used to train models, and you're not financially supporting an industry that many people find questionable. Also, if you want to, you can use "abliterated" versions which strip away the censoring that labs do to cause their models to refuse to answer certain questions, or you can use fine-tunes that adapt it for various other purposes, like improving certain coding abilities, making it better for roleplay, etc.
You don't need that much VRAM to run the very largest models, these are MoE models where only a small fraction is being computed with at any given time. If you plan to run with multiple GPUs and have enough PCIe lanes (such as with a proper HEDT platform) CPU-GPU transfers start to become a bit less painful. More importantly, streaming weights from disk becomes feasible, which lets you save on expensive RAM. The big labs only avoid this because it costs power at scale compared to keeping weights in DRAM, but that aside it's quite sound.
While you can run with weights in RAM or even disk, it gets a lot slower; even though on any given token a fraction of the weights are used, that can change with each token, so there is a lot of traffic to transfer weights to the GPU, which is a lot slower than if it's directly in GPU RAM. And even more slower if you stream from disk. Possible, yes, and maybe OK for some purposes, but you might find it painfully slow.
Why do you expect that a weighted random text generator will ever behave in predictable way?
How can people be so naive as to run something like Claude anywhere other than in a strictly locked down sandbox that has no access to anything but the single git repo they are working on (and certainly no creds to push code)?
This is absolutely insane behavior that you would give Claude access to your GitHub creds. What happens when it sees a prompt injection attack somewhere and exfiltrates all of your creds or wipes out all of your repos?
I can't believe how far people have fallen for this "AI" mania. You are giving a stochastic model that is easily misdirected the keys to all of your productive work.
I can understand the appeal to a degree, that it can seem to do useful work sometimes.
But even so, you can't trust it with anything, not running it in a locked down container that has no access to anything but a Git repo which has all important history stored elsewhere seems crazy.
Shouting harder and harder at the statistical model might give you a higher probability of avoiding the bad behavior, but no guarantee; actually lock down your random text generator properly if you want to avoid it causing you problems.
And of course, given that you've seen how hard it is to get it follow these instructions properly, you are reviewing every line of output code thoroughly, right? Because you can't trust that either.
> How can people be so naive as to run something like Claude anywhere other than in a strictly locked down sandbox that has no access to anything but the single git repo they are working on (and certainly no creds to push code)?
> This is absolutely insane behavior that you would give Claude access to your GitHub creds. What happens when it sees a prompt injection attack somewhere and exfiltrates all of your creds or wipes out all of your repos?
I don’t understand why people are so chill about doing this. I have AI running on a dedicated machine which has absolutely no access to any of my own accounts/data. I want that stuff hardware isolated. The AI pushes up work to a self-hosted Gitea instance using a low-permission account. This setup is also nice because I can determine provenance of changes easily.
> How can people be so naive as to run something like Claude anywhere other than in a strictly locked down sandbox that has no access to anything but the single git repo they are working on (and certainly no creds to push code)?
Because it’s insanely useful when you give it access, that’s why. They can do way more tasks than just write code. They can make changes to the system, setup and configure routers and network gear, probe all the iot devices in the network, set up dns, you name it—anything that is text or has a cli is fair game.
The models absolutely make catastrophic fuckups though and that is why we’ll have to both better train the models and put non-annoying safeguards in front of them.
Running them in isolated computers that are fully air gapped, require approval for all reads and writes, and can only operate inside directories named after colors of the rainbow is not a useful suggestion. I want my cake and I want to eat it too. It’s far to useful to give these tools some real access.
It doesn’t make me naive or stupid to hand the keys over to the robot. I know full well what I’m getting myself into and the possible consequences of my actions. And I have been burned but I keep coming back because these tools keep getting better and they keep doing more and more useful things for me. I’m an early adopter for sure…
Well, one of the other reasons I suggest running it in a strictly limited container is that you can then run it in yolo mode.
In fact, I use the pi agent, which doesn't have command sandboxing, it's always in yolo mode, I just run it in a container and then I get the benefit of not having to confirm every command, while strictly controlling what I share with it from the beginning of the session.
> How can people be so naive as to run something like Claude anywhere other than in a strictly locked down sandbox that has no access to anything but the single git repo they are working on (and certainly no creds to push code)?
Because it is much easier to do and failure rate is quite low.
Who would have guessed that running a binary blob dev tool, that is tied to a SaaS product, which was mostly vibe-coded, could lead to mysterious, hard to debug problems?
All traffic is monitored, all signal sources are eventually incorporated into the training set in one way or another. The person you're responding to is correct, even a single API call to any AI provider is sufficient to discount future results from the same provider.
ok! So if someone uses an existing, checkpointed, open source model then the answer is yes the results are valid and it doesn't matter that the tests are public.
You live in a conspiracy world. Those AI providers don't update the models that fast. You can try ask them solve ARC-AGI-3 without harness and see them struggle as yesterday yourself.
And I've been using this commonly as a test when changing various parameters, so I've run it several times, these models get it consistently right. Amazing that Opus 4.7 whiffs it, these models are a couple of orders of magnitude smaller, at least if the rumors of the size of Opus are true.