Hacker Newsnew | past | comments | ask | show | jobs | submit | syntaxing's commentslogin

I’m on a strix halo @ GPU-5 with MTP and I get 600 prefill and 30 TG which pushes it into a very usable range. The odd thing is that Dflash2 is really slow for me, like sub 10 TG.

Vulkan or ROCm backend?

I've also got a strix halo box. 30tok/s would we usable, but I wonder how it compares to the Qwen3.8-Flash-next - I get about 40tok/s running that on Halogen and it feels like using Claude 4.6.


Vulkan, I have never used ROCm on it but have been debating since the latest big update. How is your prefill? Do you hit over 1K? If it’s 1000K prefill, and 40 TG, I might have to try this over the weekend. Also, can you fit 128K without offload the ngram onto SSD?

I have not measured pre-fill, but it's said to be around 1000. It feels very snappy and unlike my experience with running 27B models the performance stays pretty flat even as the context increases. Unfortunately, we don't know how Halogen is doing this because it's closed source, but I think AMD should offer that guy some $$$ because he's done a lot of good work getting more performance out of Strix Halo.



Thanks! I really like how the author packaged everything into a container. Definitely going to give it a go over the weekend!

I've found the opposite on my R9700 (n-max=7, no other speculative decoding like ngram-mod, which I found slows it down). I think it depends whether your workload and system are bandwidth limited or compute limited. I see draft acceptance around 0.55, so 0.55 * 7 = 3.8 tokens per pass, which on my bandwidth-limited card takes me from 30tps to a peak of 80tps on llama.cpp (MTP peaked at ~65tps). I'm also running a Qwen fine tune whose speculative execution is better than the base model.

Strix Halo has lower compute than the R9700 but the RAM is also slower, so not sure what would be the ultimate limiting factor.


Can you point me towards the model you use, both the main model and the flash model? Curious if I can get ~30 with a higher quant.

Model, Q4_K_M: https://huggingface.co/agentionai/Signal-3.8-27B-GGUF

DFlash2, Q8_0, --spec-draft-n-max=7: https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2-GGUF

I run llama.cpp with -ctv=8, -ctk=q4. Vulkan has better throughput if you're doing single-stream decode, but ROCm has better throughput if you have "--parallel 2" or higher. If supporting parallelism, unified kv cache should be off, especially with Vulkan.

Of course, some of these may be specific to my card so try variations for your hardware. Hermes can concoct a test suite and run some tests for different llama.cpp parameter permutations to find something optimal.


Thanks! Have you seen issues with quantizing the kv cache?

> audio-visual performance close to Gemini 3.8 Flash and overall audio performance that exceeds Gemini 3.8 Flash

Wow crazy if true. I think Gemini's audio capability and multi language was the "selling point" for a lot of people. Other capability also matches or exceeds 3.8 Flash.

They also made a new harness but github link seems to 404.


probably their distill target

I said this before but I wonder if Dan Kan will reboot Atrium. Rally up some old partners and hope Anthropic buys them out for a couple billion.

It’s not obvious but you can use this with your own local (or any) models.

https://support.mozilla.org/en-US/kb/smart-window-byom


I tried adding my own endpoint, but the save button was still greyed out so it did not work. Maybe the feature is not fully implemented yet? In about:config, I could add the endpoint and api key, but not the model name.

Found the reason, HTTP-endpoints has to be localhost or 127.0.0.1, if you want to use other endpoints it needs to be HTTPS.

Keys that needs to be set in about:config

- browser.smartwindow.apiKey

- browser.smartwindow.customEndpoint

- browser.smartwindow.model

Adding my http endpoint (using LAN IP) seems to work, getting around the check in the regular settings.


This should BE front and centre and what is sold, not ads for a 3rd party.

I would pay a good chunk of money if Apple released a local AI hub to coordinate all AI usage locally (including photo indexing).

That is rumored to exist.

https://www.macrumors.com/guide/apple-command-center/

Mark Gurman seems to get extraordinarily accurate and detailed leaks. I wonder how Apple got this leaky. Pretty much everything about the Duo was known months in advance.


Well, either corporate security got incompetent or they intentionally let select leaks exist to "test the market response". Established companies tend to get more risk averse overtime.

Photo indexing is already local. What specifically are you looking for?

You should look at Osaurus.

Wow thanks for the link. I have zoom on my personal laptop which isnt ideal. I always wanted to run it sandboxed

Has anyone have good success using AI generated CAD parts? I’ve been trying but it’s always 95% there, but with all hardware, you need 100% right. It’s often quicker and cheaper for me to do it by hand (but I was a mechanical design engineer for about a decade prior)

Just this morning I used OpenSCAD for the first time. I got Gemini (just the chatbot, no harness or MCP) to help me design a water-bowl holder for our senior dog (she kept knocking it over).

It was probably an easy model to get right, since the only "critical" dimension was the radius of the interior. But I was able to tweak the numbers on the support length & some other details relatively easily, and I had a working solution 2 hours of print time later.

(I had a CAD class in highschool, but haven't used it since).

I suspect this may be another case of "LLMs are mainly good at things you're bad at."


assemblies and tolerances are where AI generation quickly gets hard. but, for simpler things, it can be surprisingly good, especially for people who has no CAD knowledge. it certainly might feel a bit like brute-forcing through the LLM - but this also applies to vibe-coding..

A screw can't be 99.9 percent correct—it has to be 100 percent correct. Microns matter.

Claude Code: Oops, I am sorry but I was wrong.

As you say it is quicker to do it by hand. Now.

I certainly wouldn’t entrust this job to a general-purpose (universal) tool that gets it 99.9 percent right.

But a domain-specific harness? Maybe it could work.

Hardware and engineering dont tolerate hallucinations, trust-me-bro numbers, deviations.


Surprised no one is talking about it but the 0.1 version bumped the parameters from 284B to 552B but “more efficient”, particularly kv cache usage

I wonder if that’s why 3.8 got so much better? Mixing the reasoning traces from both sides seems to be effective.

I’m surprised they allow open lid drinks in the lab. One wrong bump and poof 300K easy.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: