The numbers, and how they were measured
Every number on this site comes from a test on real hardware, written down the day it ran. They are from one machine and from development versions, so treat them as a direction, not a promise for your card.
- GPU
- NVIDIA RTX 5090, 32 GB, PCIe 5.0 x16
- Memory
- 64 GB RAM
- Storage
- NVMe SSD (PCIe 4.0)
- Versions
- elsieOS 0.1.10-0.1.16, November 2026
- Compared with
- Ollama 0.34 on the same machine
The two chat models: Mistral Small 3.2 24B (Q4_K_M, 13.3 GiB of weights, 16k context, kept loaded) and Qwen3.8 27B (q5_K, 18.2 GiB, 64k context, with a vision part). Together they do not fit on the card, so every switch below is a real one.
Requests were sent from another computer over the network, at temperature 0, and timed from sending to the last word of the answer.
Switching between two chat models
One-line requests that alternate between the two models. Each time includes the switch and a short answer (about 0.1-0.2 s of it). Medians of 4-5 switches.
| Request | Ollama, from RAM | Ollama, from SSD | elsieOS 0.1.11 (stop + load) | elsieOS 0.1.14 (park + wake) |
|---|---|---|---|---|
| Qwen → Mistral | 1.6-3 s | 5-11 s | 1.4 s | 0.87 s |
| Mistral → Qwen | 1.6-3 s | 5-11 s | 1.9 s | 1.01 s |
| Qwen → Mistral, ~3200-token system prompt | 2.3-3.7 s | 5.7-11.7 s | - | 0.89 s |
| Mistral → Qwen, same prompt | 2.3-3.7 s | 5.7-11.7 s | - | 1.03 s |
0.1.11 numbers are with the model files already in the page cache; read from the SSD they were 7.5 s and 8 s. With a long system prompt Ollama also reads the whole prompt again after a switch (~0.65 s); elsieOS keeps the prompt cache.
Where the time goes
| Step | Mistral 24B | Qwen 27B |
|---|---|---|
| Park (weights to pinned RAM) | 0.26-0.29 s | 0.36-0.38 s |
| Wake (weights back, new context) | 0.31-0.32 s | 0.51-0.53 s |
| First park after a start (setting up the pinned RAM once) | 1.76 s | 2.38 s |
| First load from the model file after a start | 7.5 s | 8.0 s |
The copies run at ~50 GB/s, the limit of the bus. Qwen's wake also reloads its vision part.
Sharing the card with ComfyUI
ComfyUI 0.19.1 with two everyday workflows, next to Mistral Small 3.2 kept loaded (~16.8 GiB on the card with its context). Neither workflow fits beside it, so every hand-over is real.
- Image
- FLUX.2 [klein] 4B, 1024x1024, 4 steps, peak ~17.1 GiB
- Video
- Wan 2.2 image-to-video A14B fp8, 480x480, 49 frames (3 s at 16 fps), peak ~31.1 GiB
| What happened | elsieOS 0.1.16 |
|---|---|
| Image job while the chat model is idle | chat model parked, job started after 0.28 s; ran 3.7 s |
| Chat right after (ComfyUI idle but holding its models) | answered in 0.64 s: ComfyUI dropped its models (0.3 s), the chat model woke (0.31 s) |
| The same image job again | started after 0.27 s, ran 2.4 s |
| Chat sent during 3 queued image jobs | waited, answered 0.86 s after the last job |
| Chat sent 20 s into a 66 s video job | waited, answered 0.63 s after the video ended |
| The kept-loaded model coming back by itself | back 4.8 s after the job ended |
No out-of-memory errors; every job ran fully on the card (ComfyUI's log shows no offloading).
Breaking an update on purpose
On the real machine, with Secure Boot on: a normal update, then a version built to fail its health check, then a frozen system.
| Test | Result |
|---|---|
| Update 0.1.0 → 0.1.1 | written to the spare slot; the trial start passed and the version was kept. Start-up time 29 s. |
| Update to a deliberately broken 0.1.2 | three trial starts about 36 s apart, each failing its health check; the fourth start went back to 0.1.1 by itself, about 2.5 minutes after the first, with nobody at the machine |
| System frozen on purpose | the hardware watchdog restarted the machine about 30 s later; it answered again 56 s after the freeze |
The same tests passed first in a virtual machine with Secure Boot.
15.5 hours as an Ollama replacement
elsieOS 0.1.10 took over a working Ollama machine with the same network address and keys, so none of the apps calling it was changed: a chat app, a coding assistant and a laptop.
Chats
from real apps, over 15.5 hours
Answered
the one failure was a GPU-sharing bug, fixed in 0.1.11
Median, chat app
kept-loaded model, 23 chats
Back to the old system
every service running again after the switch back
The network cable was also pulled three times on purpose (29-49 s): nothing on elsieOS failed, and requests sent meanwhile arrived once the link was back.
Caveats
- One machine, one card. Slower PCIe links, less RAM or a smaller card will change every number here.
- Development versions. Numbers are re-measured as elsieOS changes; this page shows the latest.
- Ollama was measured as configured on that machine (flash attention, q8_0 cache, 2 parallel requests).
- Answer speed itself is the card's: about 100 tokens/s for the 24B model on both systems. elsieOS makes switching fast, not answers.