Qwen3.8 · One RTX PRO 6000 project homepage · Documentation

Single-GPU inference reports, reproducibility artifacts, and experimental inspection tools.

Qwen3.8 on a single RTX PRO 6000

Independent deployment reports, reproducibility artifacts, and benchmark receipts for running large Qwen3.8 models and related inference experiments on one NVIDIA RTX PRO 6000 Blackwell Workstation Edition.

November 3 — 262K production context

The production default now configures eight 262,144-token sessions with resident INT8 KV, all 24,576 GPU experts, matched vision and 8,192-token prefill chunks. Seven normal production canaries passed, including a 259,999-token request and its repeated prefix. Full eight-way deep generation remains unqualified; the operator explicitly approved this default after those test limits were reported. Results, deployment and full test data. A later credential update is approved and live; fresh recovery checks passed. [Recovery…