Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!


New on LowEndTalk? Please Register and read our Community Rules.

All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.

Kimsufi/Soyoustart/OVH Rise New Price

1…734735736737738740»

Comments

  • mrclownmrclown Member

    @Neoon said:

    @Mainfrezzer said:

    @Neoon said:
    Actually the KS-LE-6 is amaze.

    The Motherboard does support 8 Memory channels.
    Since we are stuck at 128gigs, is just 4 but....
    LLM's seem to kinda enjoy that bandwidth.

    Usually after like 20k context the tps on a KS-LE-B goes down drastically
    Here though 60k hold steady, I wonder if 128k would work as well.

    what do you have running on it atm and how quick are we talking?

    Nobody talks about quick.
    Its about keeping the initial tokens per second even on long context window.

    Which makes it interesting, for long, agentic workloads.
    Then it doesn't really matter if it takes longer right, at least for me.

    I am running Qwen 3.8 Flash Next, newest unsloth 3bit quant.
    I try 4bit later also 128k context.

    3.8 Tokens per second, slow as fuck but stable.

    Good for experiment but not practical. 20$ subscription for Claude, Gpt or open router might worth way more than using LLM on KS-LE-6.

    Thanked by 1allthemtings
  • NeoonNeoon Community Contributor, Veteran
    edited 7:14PM

    @mrclown said:

    @Neoon said:

    @Mainfrezzer said:

    @Neoon said:
    Actually the KS-LE-6 is amaze.

    The Motherboard does support 8 Memory channels.
    Since we are stuck at 128gigs, is just 4 but....
    LLM's seem to kinda enjoy that bandwidth.

    Usually after like 20k context the tps on a KS-LE-B goes down drastically
    Here though 60k hold steady, I wonder if 128k would work as well.

    what do you have running on it atm and how quick are we talking?

    Nobody talks about quick.
    Its about keeping the initial tokens per second even on long context window.

    Which makes it interesting, for long, agentic workloads.
    Then it doesn't really matter if it takes longer right, at least for me.

    I am running Qwen 3.8 Flash Next, newest unsloth 3bit quant.
    I try 4bit later also 128k context.

    3.8 Tokens per second, slow as fuck but stable.

    Good for experiment but not practical. 20$ subscription for Claude, Gpt or open router might worth way more than using LLM on KS-LE-6.

    Basically, to get something usable like 30t/s you would need a GPU.
    Maybe the 1bit quant would be faster but to much brain damage.

    I just found this out by accident, I have expected speeds at about 4t/s but I didn't expect the tps to be stable.

  • tucibeyintucibeyin Member

    @allthemtings said:

    @tucibeyin said:
    The product is being delivered

    Which?

    6 hybrid

  • plumbergplumberg Veteran, Megathread Squad

    KS-LE-3 in stonks

  • op23op23 Member

    It’s nice to see that you can get stable 3.5 tps on a LE-6.
    Sure it’s no coding assistant but this doesn’t matter for overnight tasks. I could just throw something on it and come back next day to find what I needed. Perfect and I don’t need to give up a kidney to buy hardware to do the same.

  • NeoonNeoon Community Contributor, Veteran

    @op23 said:
    It’s nice to see that you can get stable 3.5 tps on a LE-6.
    Sure it’s no coding assistant but this doesn’t matter for overnight tasks. I could just throw something on it and come back next day to find what I needed. Perfect and I don’t need to give up a kidney to buy hardware to do the same.

    I use a modified python script, with a one shoot prompt.
    It keeps doing the task until it deems to be done.

    It runs at a stable 3.6 to 3.8, at 60k context right now.
    Currently I told it to clone a certain browser game.
    https://stats.neoon.net/monitor/f453e94d2c671cadd6c7eba1cca35a0b

    No dent in CPU usage more or less, perfection.

  • webcraftwebcraft Veteran

    Anybody got something in LIM/SBG ready to trade?

  • barbarosbarbaros Member

    My 100 euro offer still stands

  • @barbaros said:
    My 100 euro offer still stands

    double it and give it to the next person

Sign In or Register to comment.