That’s only true if you think AI is the only reason to own a powerful and efficient server. Mine does plenty of traditional server stuff too.

I can do traditional server stuff on any old computer with a big hard disk and a decent amount of RAM. That's not worth $3500-$4000. When RAMpocalypse is over and we can buy a Strix Halo for under $2000 again, the math starts mathing. It becomes a pretty great desktop computer that also happens to run AI pretty well at a pretty good price.

Yeah, but that computer can’t also do the AI stuff. And not everybody has a desktop with multiple 32GB GPUs available.

I’ll admit though I’m biased because I bought my board for $1600 back before the prices went crazy.

Oh no a tough constraint that will lead to further innovation like deepseek. How terrible.

I spent around 5k on a server for "AI stuff" and it's currently doing no AI, because local LLMs (at least on systems with 32 GB VRAM) can only do only very basic stuff; this includes Qwen3.8 - in spite of the reverse engineering blog post, when I've tried Qwen to do a similar task, it flunked miserably. Additionally, I've read on some informal sources, the next step in quality is at 256 GB, not 128, which is very expensive (it's around 10k). 10k for privacy is... a toy for rich tinkerers, considering that most the people have their email on the cloud.

Same, I've had 96GB VRAM available for more than a year, and testing all the local LLMs as they become available, and read about people using models with Q2 quantization saying they get real work done, meanwhile I can't run 4x the size models with BF16 and get them to not eventually fuck up something.

Currently, I think most models are somewhat benchmaxxed, as my own benchmarks and others I know who also keep private benchmarks, don't show nearly as much needle movement as the press releases tend to want you to believe.

I can't wait for local LLMs to mature enough so I can move everything locally, but we're sadly not there yet today.

If you can't get Qwen 3.8 27b to work well that's a skill issue.

You just suck at using ai

Which exact quantization you use of that particular model?

I've just spent two days testing it qualitatively (after the quantitative results were kind of bad, but promising), the official BF16 weights, and it keeps fucking up very basic things, like doing "rm -rf sessions/$id" without really checking or making sure $id is defined, so it ends up deleting more.

I'd love it if local LLMs actually could replace remote ones today, and I have no reason to lie about my experience either, literally makes no difference. But I too got hopefully reading the sentiment on the internet about Qwen 3.8, but it's just not there (yet) for local professional software development.

Verbatim logs from an open session where this literally just happened:

    Wait, let me reconsider. Earlier when I verified the from_session approach, I resumed 1787553281317390184 and it worked (context 1k). Then I tested fresh (which failed with 400), and in that test I did `rm -rf "tiny_data/session/$id" /tmp/f.txt` where $id was the fresh session's id (empty, since it failed). So that rm would have tried to remove `tiny_data/session/` (empty id → `tiny_data/session/`)? Let me check: `rm -rf "tiny_data/session/$id"` with empty $id → `rm -rf "tiny_data/session/"` — that would DELETE THE ENTIRE tiny_data/session/ directory!!!
    
    Oh no. That's what happened. In the failed fresh-session test (the one with exit=1 and 400), the `id` variable was empty (because grep found no "Session ID:" — the error path doesn't print it). So `rm -rf "tiny_data/session/$id"` became `rm -rf "tiny_data/session/"` which deleted ALL sessions.

Oh no indeed :)

To be fair to Qwen, we've all been there.

Maybe I got lucky (as I know others have had this issue), or it recovered by itself, but never had this happen with GPT models for as long as I've used codex, and I run it with "--sandbox danger-full-access --dangerously-bypass-approvals-and-sandbox" 100% of the times.

No, I mean, every human who spends enough time working with computers eventually makes a dumb blunder, very often this specific one (using `rm -rf` in a way that has further reaching effects than intended). But, I do want to use models that don't make those kinds of mistakes, which is why I happily pay for a bunch of subscriptions for the best models.

I don't disagree with you - I definitely "suck at using ai", which means: I don't have tons of time to waste, to obtain a result that it's both unstable (tweaking for a certain class of tasks may make another class of tasks worse) and useless (because anyway I have to use cloud AI for harder tasks). But it's certainly useful in winter to keep the room warm.

With 32gb VRAM you should be able to run Qwen3.8-27B at either Q5 with 262k context, or Q6 with 200k context. Additionally you probably want Retrieval Augmented Generation (RAG) either in your harness or the server (e.g., in Open WebUI). Admittedly I've only enabled web search in my harness (OpenCode supports the Exa search API by default) and use straight llama-server.

As far as next step in quality being 256gb, that's largely correct. Qwen 27B is currently the best model for 32gb VRAM, and you don't get better models at a reasonable speed until 256gb.

Yes, I've tested it at Q6, and it was terrible at the task.

An "AI" server can do traditional server stuff but a traditional server can't do AI stuff (inference)

how much performance (tok/s) can you expect from 128GB Strix Halo? assuming this model will be released with FP8

also can you use it for fine tuning?

The Strix Halo and DGX Spark are pretty danged slow, relatively speaking. I don't recall exact numbers, but with MoE models in this size ballpark (Laguna S 2.1), I seem to recall I was seeing about 20-25 t/s with a big context, which is close to usable. Qwen 3.8 27B crawls on this hardware, though, at 10-16 t/s, definitely not comfortable for interactive use. (Though this makes it seem like you can cook pretty good with a 4-bit ROCmFP4 quantization: https://github.com/julianmb/q38rocm the model does get notably dumber below six bits.)

A model similar in size to Laguna S 2.1, but with only 6B active parameters, should be a notable amount faster, so I would imagine 25-30 t/s would be a reasonable guess for where Qwen 3.8 Flash Next will land.

DFlash2 might improve all these numbers. It wasn't available last I was testing new models on the Strix Halo; I've only used MTP (which doesn't generally improve MoE models, but I believe DFlash2 can).

Given software improvements, I'm hopeful an MoE in this size range will be the sweet spot that pushes past 40 t/s and is also smart enough for real work. Qwen 3.8 27B is finally a self-hostable model that's smart enough, but it thinks so hard it still isn't really useful for agentic interactive use.

On the DGX I get 44.5 tokens per second (NVFP4). With 8 concurrent it's 241 t/s total.

I am using the PrismaAQUA

standard 9.7 t/s

+ Dflash2 30 t/s

+ torch-compile 37 t/s

c8 = 177 t/s

What model? Also, I don't know what "the PrismaAQUA" means, ddg thinks it's a CPAP machine, which seems unlikely to help with inference performance.

Also, 4-bit has measurable intelligence loss. Sometimes worth it, but, at this size models are barely smart enough at 8 or 6.

Qwen3.8-27B-PrismaAQUA-5.5bit-vllm

The output quality is higher. It's held at full precision (not quantized).

You can only use up to 90gb for the GPU, so it doesn't fit

No. In Linux, you can ignore the BIOS (well, set it to 512MB) and set GTT to allow almost the entirety of memory for GPU. If you don't run a desktop GUI (and why would you), you don't need memory for much of anything else. Just a few GB for the OS and llama.cpp.

These are roughly the settings I use: https://github.com/kyuz0/amd-strix-halo-toolboxes#kernel-par...

A 3060ti 8gb, released in 2020, has 448 GB/s of bandwidth compared to the Halo 256 GB/s

The 3080ti is 912.4 GB/s

And the newly announced/launched Apple M6 has 170GB/s of unified memory bandwidth, meanwhile M5 Ultra gets 1.2TB/s of unified memory bandwidth. https://www.apple.com/newsroom/2026/08/apple-introduces-m6-a... Not sure if the first one is a typo on their press release, can't be just 170GB/s then be pushed for AI use, can it? Could be a different measurement I suppose...

You got me curious so I looked up the previous chips[0].

Memory bandwidth M1: 68 GB/s M2: 100 GB/s (47% increase) M3: 100 GB/s (0% increase) M4: 120 GB/s (20% increase) M5: 153 GB/s (27.5% increase)

So, M6: 170 GB/s (11% increase) doesn’t seem impossible, though I would have expected more.

[0]: https://www.jdhodges.com/blog/apple-cpu-compared-m1-m3-m3-m4...

The different models of chips and memory config have very different memory speeds as well.

Eg the M4 Max 128GB has a bandwidth speed of 500GB/s+. And that's true for other models as well.

But as you note, the base speed has also increased over the versions.

From a bandwidth perspective, the ultra is like 8 M5s fused together (@ 150GB/s), that's how it gets to the 1200.

Historically the Pro doubles the base, the Max doubles the Pro, and the Ultra doubles the Max.

If an M6 ultra were released today it would be 1.36TB/s.

does that mean they're measuring bandwidth differently than how others (like nvidia) does it? memory bandwidth is the gating factor of running models locally, so if it's actually 8x 150GB/s, it may help something like prefill, but would it actually speed up decode comparatively?

No, they use the same definition of memory bandwidth as others, but Apple Silicon has a lot of memory channels. In previous generations, prefill has been compute-limited and decode is fast.

https://blog.exolabs.net/nvidia-dgx-spark/ outlines a combination of a DGX Spark and an M3 Ultra that took advantage of fast prefill on the Nvidia hardware and fast decode on Apple Silicon.

GFX vram is still faster.

Not a hardware engineer, but it's mainly because of RAM wires/channels (not implying that this is "simple" form an engineering perspective).

Using the published bandwidths, the math is 170 * 1 and 153 * 8.

But 170GB/s is almost nothing? None of the RTX 50 series GPUs has that low bandwidth, you have to go back two generations of nvidia GPUs to get closer to that, and then it's the cheapest of the series, RTX 3050, which has ~170GB/s.

Even the GTX 1080, launched ten years ago, has double the bandwidth!

This must be some different way of measuring the bandwidth right? Since they explicitly say this for AI, but the numbers they share don't show that at all. Or I gravely misunderstand something here.

> Since they explicitly say this for AI

That's marketing spin indeed (or lies, if you prefer).

But it also has 8GB of RAM.

Enjoy Gemma 4 E2B at blistering speeds, I guess?

My point was more: this was the 2nd lowest end card from a generation 6 years ago, and it had way higher bandwidth than today's alleged flagship.

Small amounts of fast ram vs huge amounts of slower ram.

It costs more than the strix to just buy regular ddr5 ram sticks today.

Arguing with people on here that you should buy GFX cards instead of overpriced Macs for inference is a lost cause. Either Apple astroturfs this forum hard enough to convince people Macs are good for local llm, or people are REALLY stupid and don't understand how local inference works and think that the dogshit slow 40-50 tok/sec is standard.

I have had an impossible time getting 120B or better models running on Strix Halo (especially under Windows) with any large context windows. And 30-40 tokens/second is fine, but not the fastest.

For the most part lately I have been sticking with Qwen 3.8 27b and that thing will easily suck up 64gb of ram. Add in docker with some additional programs running and it's really easy to eat up 128gb of ram.

I found a couple of different 4-bit quantizations of Laguna S 2.1 that run pretty well with pretty big context (also quantized, to 8 bits, I think). Unfortunately, Laguna isn't better than Qwen 3.8 27B, which I'm able to run at roughly the same speed on my desktop machine, so I don't use Laguna or the Strix Halo very much, lately. (It's also too hot for me to be running heaters for inference. It's been ~110F most days for the past few weeks.)