the hardware needed really depends on the model size and quantization. High memory systems like the Gorgon Halo could run some very capable local models but matching something like Opus in coding quality and long-context reasoning is probably still more about model architecture and training than raw hardware. For most people, a strong local model that runs comfortably is probably a better choice than trying to brute-force the very largest models
I believe the right way to think about this is not to ask for a local model to be good at everything, but rather to finetune smaller local models, that fit your hardware, to do specific tasks very well.
You can then get very close to online model capabilities, but just at the tasks you trained your models for. If you do not have unlimited hardware resources – and who does really – I strongly believe it is the way to go.
I am finetuning qwen3.8 27B as a builder role for my projects on an intel B70. I am also considering, after the finetune proves its worth, to get a 128G strix halo box for a reviewer model – again finetuned.
If you follow this train of thought, your first step is not to think about hardware at all, but rather to understand what kind of data you can leverage from your own projects for the finetuning process, as this is the only step not automated for you – you could use curated datasets, but models exist that are already trained on these, so what would be the point?
Realistically you’re probably not matching the likes of Opus on anything local that’ll run in your money or power budget. Maybe Sonnet or Haiku is a better target and then have a delegation model where complex tasks still go to the hyperscalers.
I wish there was an industry normalised LLM utility score. One that captures when a model gets generally useful and accurate at that pinned level. The level is versioned and can be adjusted over time. Then the race I’m most interested in is that bang/buck race towards zero.
I think you can get good-enough performance on routine tasks with script help (you’d need opus to babysit and effectively tool the local model). This is great for things like keeping worktree rebased or doing CI/CD monitoring (which is super duper expensive if you try to have frontier do it, the CI token budget quickly swamps development).
I think the interesting question is less “what hardware matches Opus 5.5?” and more “what local setup gets close enough for the tasks I actually care about?”
The hardware requirements for true frontier-level performance are still pretty wild, but local models are improving quickly. For coding especially, a smaller model with a good agent harness, tools, and the ability to iterate can be much more useful than benchmark numbers suggest. The trade-off is usually that you spend more time supervising and validating the work.
For me, the real tipping point would be when a local model can handle long-running tasks reliably enough that the extra inference time and occasional intervention are cheaper than a frontier subscription. We’re getting closer, but I don’t think we’re quite there for the hardest coding workloads yet.