[SwellJoe]

Yep, I have a Strix Halo and while it can run models bigger than Qwen 3.6 27b, it's not usable interactively when you do. ds4 patched for ROCm works, but at such a slow speed, it's not usable for coding agents.

The Nvidia boxes have only slightly more memory bandwidth, so I wouldn't expect them to be notably faster. At least not enough to make it useful interactively at that scale.

[zozbot234]

Why does everyone expect interactivity from local AI? It's not the best use of the hardware, especially not miniPC hardware. Long-term batched inference with larger and more capable models is much more feasible AIUI.

[int_19h]

I can't speak for others but IMO the only reason to run models locally right now is privacy - i.e. you don't trust any of the cloud providers to not look at your prompts. Price-wise the market is extremely competitive and cheap model serving favors large scale so anything that can be run locally can be run cheaper in the cloud. But if privacy is important, then it's important for everything, including traditional chatbot applications, which kinda do require interactivity.

[SwellJoe]

Even batched it's uncomfortably slow. I started to benchmark ds4 with my security vulnerability benchmark (after Qwen 3.6 dense and MoE and a bunch of cloud models), but it was going to tie up the Strix Halo for more than a day, so I decided not to run it as it would prevent me from doing other stuff with it during that time.

Even batched usage needs to be fast enough to deliver results in a reasonable time. Overnight runs are useful, 24 hour runs are...less so.

Anyway, most of the time people are talking about interactive use, and there's currently an upper bound on how large a model can be for local hosting on a reasonable budget (i.e. not a crazy amount more expensive than what a high end developer desktop or laptop costs). The sweet spot is probably currently the big Qwen 3.6 or Gemma 4 models, which are in the ~60GB range for 8-bit quantization plus a large context.

[hedgehog]

The 6-bit versions + 8-bit KV cache seems to save a good bit of memory without a significant loss of quality. The Qwen 35B is pretty fast in my testing, but MiniMax M2.7 230B is in some ways faster (way fewer tokens to arrive at an answer) even though it is much larger.

[SwellJoe]

Qwen 3.6 35B-A3B with MTP at 8 bits is blazing fast, something like 50-60 tokens per second. That's plenty fast for interactive use, so I haven't tried lower bits. Unfortunately the MoE is notably dumber than the dense model (for the case I have data about...I've been benchmarking models for security vulnerability scanning, and 27B is notably better on hard bugs).

The dense model is almost usable, but feels really sluggish, even with MTP. I think it's about 12-15 tokens/second on the Strix Halo. Slow enough to where I'd rather pay to use a cloud model.

I might try the 6-bit version of the dense model to see how it behaves, though. Maybe it'll retain its bug hunting abilities while making it fast enough to use interactively and not take all day for benchmark runs.

[hedgehog]

Same chip, with a 6 bit 35B and 8 bit KV cache I see about 500 prefill and 55 decode at 30k into the context window. MiniMax seemed a bit lower token rate but much, much less prone to 40k tokens of monologue before generating an answer. A pattern I like is to use a smaller model to do most execution and then a larger model to review transcripts and output and do any fixups and tooling improvements (this is all batch jobs so all I care about is overall throughput).

[milch]

What hardware do you need to run MiniMax M2.7 230B locally?

[hedgehog]

Ryzen 395 is what I'm using, anything with 128GB+ of RAM accessible to the GPU should work fine for a 4 bit version of the model (so Spark or Mac Studio should be ok too).