Hacker News

points by layoric 2 days ago | 0 comments

IMO then it is missing the forest for the trees regarding open weight models. Qwen 3.8 Flash Next or 27B, for example, are accessible enough to run at home/office at very usable speeds. Provided memory companies don't collude to keep margins at 80-90% they are at the moment, it seems to me this will become a lot more accessible for many use cases while models at this size still have room to improve.

[–] jimbokun 2 days ago

Except it’s very difficult for home users to afford the GPUs needed to run them with acceptable performance.

[–] layoric 19 hours ago

Strata, Radiance, Maya, etc have really changed that IMO. I use an old X99 setup with 2x A4000s and get a solid ~70 tk/s decode, ~2k prefill, and with 128GB 2133 DDR4, I have memory for quite a few sessions worth of cache so multiple agents can run with decent switching. Probably other performance nobs to setup, but I've found this is great with Qwen 3.8 Flash Next IQ3_S .

Yes, hardware has gotten insanely expensive, but software is moving so quickly and a lot of effort is pointed at eking out every bit of performance, and we are seeing a lot of success. Really looking forward to Qwen 4+. Yes Opus 5.5 is better than what I can use at home, but the way I use these models, I don't feel I need that. I prefer to be in the drivers seat, and these delegate the simple tedious parts to models which they are getting really good at.

[–] notenchange 2 days ago

I have real world data here. On a 5060 Ti (16GB) with 64GB RAM, I’m getting ~65 tok/s on decoding and ~450 tok/s prefil for Qwen 3.8 Flash Next. 256K context.

This is totally suitable (speed wise) for coding.

This is not some top of the line hardware (might be now, but it should be run of the mill for people that built their system <1.5 years ago).

[–] ozgrakkurt 2 days ago

How do you even fit that model in there? What quant is it? Are you using llama.cpp?

[–] notenchange 2 days ago

No, I'm using strata. This is a MoE model, so it keeps the last used model in VRAM, and the rest in RAM. If it needs, it swaps the less used model in VRAM for the model in RAM for the next token.

Seems very slow, but it's actually fine.

[–] ozgrakkurt 1 day ago

Thanks, the information in their docs are really promising. I have been meaning to try it but wasn't able to yet.

[–] notenchange 2 days ago

P.S. Sorry, where you read “models” I meant experts.