Name the Blackwell serving cell you are actually in

<!-- SC_OFF --><div class="md"><p>I tested two llama.cpp builds on the same <strong>Qwen3.6-27B-NVFP4</strong> model.</p> <p><code>llama-bench</code> reports the model label as

<p>This article compares two AWS GPU instance families for serving a small language<br /> model, using a payload that is byte-identical on both. The older family loses<br />

<!-- SC_OFF --><div class="md"><p>That foodtruck bench post showing deepseek v4 matching gpt-5.2 at 17x cheaper got me thinking. if frontier cloud models are that overpriced for

<!-- SC_OFF --><div class="md"><p>Got Qwen 3.5 27B to 1,103,941 tok/s on 12 nodes with 96 B200 GPUs.</p> <p>At that rate you process 50,000 insurance policy documents in hours

<p><em>Originally published on <a href="https://hexisteme.github.io/notes/it-fit-in-memory-and-was-still-unusable.html" rel="noopener noreferrer">hexisteme notes</a>.</em></p>

<p>FP8 gave us a clean 1.5x on Qwen3-8B serving throughput on an RTX PRO 6000 Blackwell (1,725 to 2,597 tok/s at concurrency 32, vLLM). The uncomfortable question is always the