← News
Arena

A three-3090 Flash-Next report: 334.86 tok/s needs its ESXi, patch and MTP conditions

A new community report records QEngine 1.3.5.7 on three 3090s. Power caps, missing P2P, a tensor patch and benchmark phases are essential conditions, not optional footnotes. No independent AraoAI reproduction is claimed.

Source: club-3090 社区自报及维护者评论

On November 13, 2026, a club-3090 contributor posted a three-RTX-3090 Qwen3.8-Flash-Next report. QEngine 1.3.5.7 with INT4 and MTP recorded narrative and code decode means of 259.20 and 334.86 tok/s. These are the contributor's measurements, not an AraoAI reproduction or a general speed expectation for 3090 cards.

The conditions are essential: an EPYC 7R13 system with ESXi 8.0 passthrough, a 320W cap on every GPU, no P2P and a gpu-unlock tensor patch. Power limiting and the patch act in different directions. Host-staged transfers are not equivalent to native peer access, so a comparison cannot assign the whole difference to one factor.

Boundary and soak checks do not establish preserved task ability. The issue lacks a complete quality pack and mentions a tool-call miss. A maintainer confirms that this is a rig-specific community result and flags two reporting defects: a QEngine decode-source label incorrectly names llama.cpp, and service identity is printed as question marks. Label defects and measurement validity require separate checks.

The issue and attached rig report are also different runs. The attachment was generated at 06:29:01 UTC. Its 10K phase reports paired means of 4209.96 prefill and 252.05 decode tok/s across three short-output probes. Do not combine its prefill with the issue's 334.86 code result, or represent the short probe as sustained long-output throughput. Only the attachment's paired data is submitted as an L0 draft for review.

Licensing must be versioned too. The LICENSE at tested commit 5028c50 is PolyForm Noncommercial 1.0.0 and restricts commercial use. A later repository PR says QEngine changed to MIT from 1.3.5.8; this test used 1.3.5.7. Do not project a newer license backward. Ask what makes a number valid before asking whether your machine can reproduce it.

Issue: https://github.com/noonghunna/club-3090/issues/2011 Attachment: https://gist.github.com/dgaglioni/7334f51d7f7f9070e0f302f4e80789e6 License-change context: https://github.com/noonghunna/club-3090/pull/2017