Discover · piszczek.pl

No picture published by piszczek.pl

Qwen3.8 27B at 256K: 50 TPS on a 24 GB GPU

AI Economics Tools piszczek.pl

I gave Qwen3.8's MTP drafter another 69.2 MiB of precision. Throughput fell from 50.44 to 37.02 tokens per second. That result sums up the whole experiment: the best local inference setup is rarely made from the individually "best" parts.

I wanted a dense 27B model, its full 262,144-token context, multimodal input, maximum useful quality, and speculative decoding on an NVIDIA RTX PRO 4000 Blackwell SFF with 24 GB of VRAM. The server also had to survive real agent work after printing model loaded. The experiment followed a hunch I had written about earlier: careful operation may matter as much as moving to a larger model.

More ways to use this
Open on another device
Scan this QR code to open the same public page

Scan to open this page. Your saved notes and account access are not included.

Download QR code

Open the page link

In the piece

Named in it

Keep going

Each of these is read from what this estate already kept: the feed a piece came through, the publisher behind it, the credits and topics the publisher printed. None of it is a recommendation, an identity or a record kept anywhere else.

Also from AI Economics Tools

  1. Looped Transformers Explained: How AI Thinks Longer Without Getting BiggerMichał Piszczek piszczek.pl
  2. The Average Is Free. Difference Is Not.Michał Piszczek piszczek.pl
  3. 67% more tokens per second from a tiny BlackwellMichał Piszczek piszczek.pl
  4. RSI AI Explained: What It Is and Where It Could LeadMichał Piszczek piszczek.pl
  5. Scarcity Never Dies. It Climbs.Michał Piszczek piszczek.pl
  6. DFlash Changes What Tokens per Second MeansMichał Piszczek piszczek.pl