Source: https://tensorquay.com/blog.html

THE TENSORQUAY BLOG · LAB NOTES

# Local AI. Put to the test.

What fits, what finishes, what breaks. Experiments with open models and GPUs, from our team in Singapore. The settings and the awkward results come with the good ones.

LAB NOTE 04 · LONG SESSIONS TESTED 26 & 28 SEP 2026

## [Eight long sessions. 64 GB of RAM.](https://tensorquay.com/qwen-long-sessions-rtx-5090-64gb.html)

The Qwen cache test again, on our own RTX 5090 test machine with the RAM budget of a TQ–32. Eight sessions with 32K of history averaged 33–36 tok/s per user over two runs. Why three of eight users were slower, and what happened when the cache spilled onto an SSD.

[Read the long-session results](https://tensorquay.com/qwen-long-sessions-rtx-5090-64gb.html)

8 × 32K sessions · 1 × RTX 5090 · 64 GB-class RAM
Qwen3.8-27B NVFP4 · vLLM 0.28 · 24 GiB CPU KV tier

Mean per user · run 1 / run 2 33.1 / 35.1 tok/s

Slowest 3 of 8 users about 22–23 tok/s

First repeat of each run; scale 0–50 tok/s. Synthetic prompts and 128-token replies. Not an assembled TQ–32.

LAB NOTE 02 · VIDEO GENERATION TESTED 18 SEP 2026

## [FastH3 on a 5090: time the first clip, too.](https://tensorquay.com/rtx-5090-video-generation.html)

56.6 s warm median · 5.2 s clip with audio · 4 prompts

One cold run took 147.3 seconds. Here are the samples, the settings and the memory bill. The test host had 125 GB RAM.

[Read the video generation notes](https://tensorquay.com/rtx-5090-video-generation.html)

LAB NOTE 01 · CODING AGENTS TESTED 2 SEP 2026

## [Qwen on one 5090. Keep the history in the test.](https://tensorquay.com/rtx-5090-local-ai-benchmark.html)

7.0–7.6 → 35.8–44.1 effective tok/s per session · 2 repeats each

Four scripted 32K sessions, 24 GiB of CPU cache and two repeats. Warm follow-up turns exposed a limit that short prompts missed. Plus recordings of the separate coding-agent runs.

[Read the Qwen benchmark](https://tensorquay.com/rtx-5090-local-ai-benchmark.html)

LAB NOTE 03 · MULTI-GPU SERVING TESTED 19 SEP 2026

## [Two PRO 6000s. What does GLM finish?](https://tensorquay.com/glm-5-3-flash-dual-rtx-pro-6000.html)

8 / 8 coding tasks passed · 4.5 min median

GLM-5.3-Flash, UD-IQ4_XS and a pinned llama.cpp build. Installation notes, concurrent request timings and the point where our agent workload started to struggle.

[Read the GLM setup and results](https://tensorquay.com/glm-5-3-flash-dual-rtx-pro-6000.html)

## Enough detail to question the result.

Each note separates the measured result from our interpretation, names the test configuration and explains what remains untested. Summary data is downloadable. Historical GPU test systems are described in each article; their results do not qualify an assembled workstation. Start with the [local AI guides](https://tensorquay.com/guides.html) if you are new to running models.