Turns out cheap tokens get engineers doing things that save time, like debugging and tackling Jira tickets
  • *for certain definition of self hosted.

Developed by Chinese open weights model dev Z.ai, GLM 5.1 isn’t exactly a small model. At roughly 754 billion total parameters (40 billion activated for each token generated), the model requires roughly 380 GB of GPU memory just to hold the weights and another 6-26 GB per user to maintain context depending on precision.

To serve GLM-5.3 at scale, Revolut deployed 56 Nvidia B200 GPUs, equivalent in GPU count to seven eight-GPU DGX B200 systems. But as Abitbol noted, the greatest performance gains didn’t come from adding GPUs, but rather tuning the model’s hyperparameters along with switching to NVFP4 quantization, implementing disaggregated prefill and decode, and tuning context lengths.

  • aihostbox@programming.devEnglish
    0·
    2 days ago

    The quoted numbers are the interesting part. 380 GB just for weights plus several GB of context per user means this only pays off at Revolut’s scale. What stood out to me is that the article credits the biggest gains to NVFP4 quantization, splitting prefill and decode, and context tuning rather than adding GPUs. Those tricks carry over to much smaller setups too.

    • Spite_By_Design_LLC@lemmy.dbzer0.comOPEnglish
      0·
      2 days ago

      Agreed. Economies of scale are about the only way this makes sense.

      I was just working thru the use case for purchasing a 64gb Strix (vs the typical coding I might do using OR or Open Code Go). The ROI is somewhere in the order of 7-10 years for me. So, on a purely dollar cost basis, local inference is $$$. OTOH, from a sovereignty and “own your own code” basis…there’s something to be said for it.

      Probably I will run a few tests using OR with a few outstanding tickets of mine, to kick the tyres of Qwen3.8-27B. I’ve tried Qwen3.6-35B and frankly it’s…not very good at code gen or patching.

      To boot, code gen isn’t the only thing I use LLM for, so I suppose it depends on how bloody minded I want to be about future use case (I do actually have one).

      • Womble@piefed.worldEnglish
        0·
        2 days ago

        FWIW I wouldnt recommend a 64GB strix halo for local LLM inference. I use a 128GB framework desktop for running qwen-flash-next which it just manages at Q4_K at a reasonable rake of 20-30 tk/s qwen27B is sloooow on it due to strix halo’s lower memory throughput than either a GPU or a integrated memory mac.

        Though its possible this advice will change in the next few weeks with the release of Qwen4.

  • rbn@sopuli.xyzEnglish
    0·
    2 days ago

    What I don’t unterstand about stories like this: Why aren’t the US companies making a fuckton of money then? If you can self-host something at 10% of the costs, shouldn’t that mean that a datacenter can make around 90% margin with interference?

    • Spite_By_Design_LLC@lemmy.dbzer0.comOPEnglish
      0·
      2 days ago

      In the spirit of things, I ask GLM5.3 what that set up actually looks like. Ouch.

      Fetched that page for real. Specs (from gpuworld’s DGX B200 datasheet page):

      Dimensions: 17.5" W × 87.5" D × 42.9" H (cm: 44.5 × 222.3 × 109.0)
Weight: 275 lb (124.7 kg)
Rack height: 10U equivalent
Power: up to 14.3 kW max per system

      In real-world terms:

      Height: 10U = 10 rack units = 10 × 1.75" = 17.5" tall — about the height of a bar fridge standing on its side. But that's the chassis; you'll also want ~2U of clearance for airflow/management, so plan for ~1.5U of breathing room above it in a rack.
Depth: 87.5cm = just under a metre — so it'll stick out of a standard 1m-deep server rack unless you've got deep racks (or you let the back panel hang open). Standard 42U racks are typically 1.1–1.2m deep to handle this.
Weight: 275 lb / 125kg. That's a hefty man's weight — two people minimum to lift, ideally four or a rack lift. Standard datacentre floor loading is usually fine at this density but worth checking if you're stacking multiple in a rack.

      Rack math for 7 units: 70U of just these systems — so a single 42U rack can’t hold them all (you’d need 2 racks), and at ~14kW each you’re looking at ~100kW total draw — that’s serious power/cooling, way beyond a standard office closet.

      Short version: think “fridge-sized box, half a tonne with rails, eats 14kW” — not a pizza box.