# Beelink GTR9 Pro for local LLMs: what two units measured over a month of serving

> Reviews of this box benchmark it for a week. We bought two, froze the workloads, and have been serving from them since: the software stack that works, what the numbers look like, how the two units compare, and the parts we still cannot vouch for.

- Published: 2026-08-21
- Evidence level: lab_repeated
- Funding: Self-funded internal research
- Canonical: https://agmind.ai/reports/beelink-gtr9-pro-local-llm/

Two commercially identical Beelink GTR9 Pro units — Ryzen AI Max+ 395,
128 GB of LPDDR5X-8000 unified memory, Radeon 8060S (gfx1151), dual 10GbE —
are the machines behind most of the numbers in our claim registry. This page
is what a month of using them for actual serving taught us: the software
stack that works, what the box measurably does, where the two units agreed,
and the parts we still will not vouch for.

## Why two identical boxes

Because one box cannot tell you whether a number is a property of the
hardware or of your particular unit. Buying the second one is how a
measurement stops being an anecdote: the same frozen workload, the same
pinned runtime and model artifact, run on both. Claims in our registry
carrying the `lab_unit_replicated` level are the ones where the pattern
showed up on both machines — that is what the level means, and it is the
main reason we own a duplicate instead of a second, different box.

## The software stack that works

The configuration our serving node runs today, read off the machine rather
than off a wiki:

```
kernel   7.0.0-28-generic
Vulkan   RADV (GFX1151), Mesa 25.2.8
runtime  llama.cpp server, pinned by image digest
```

The Vulkan path through RADV is what we serve on. ROCm also runs on this
hardware, and both backends passed every correctness gate we put in front of
them — the difference showed up as decode pace, not quality, and the
[side-by-side is a comparison page](https://agmind.ai/compare/llamacpp-vulkan-vs-rocm-strix-halo/).
On the networking side, both 10GbE interfaces have been up throughout; the
box's Wi-Fi is disabled on ours because a serving node has no business
holding a radio link.

The one configuration item worth reading before you buy: memory allocation
on Linux does not work the way the BIOS menu suggests, and the setting that
decides whether large models load
[is a kernel parameter](https://agmind.ai/reports/strix-halo-memory-allocation/), not a
firmware slider. That page exists because it cost us time first.

## What the box does

Serving a 35B-class MoE model at Q4_K_M, everyday requests, reasoning off:
first answer token in
**210 ms** ([strix.qwen36.interactive2.c1.ttfa-nothink](https://agmind.ai/claims/strix.qwen36.interactive2.c1.ttfa-nothink/)), decoding at
**15.9 ms/token** ([strix.qwen36q4.vulkan.c1.itl-nothink](https://agmind.ai/claims/strix.qwen36q4.vulkan.c1.itl-nothink/)). Under four
simultaneous chats the first-word wait becomes
**338 ms** ([strix.qwen36.interactive2.c4.ttfa-nothink](https://agmind.ai/claims/strix.qwen36.interactive2.c4.ttfa-nothink/)), and at eight
**865 ms** ([strix.qwen36.interactive2.c8.ttfa-nothink](https://agmind.ai/claims/strix.qwen36.interactive2.c8.ttfa-nothink/)) with
completion holding at
**100.0 % of requests** ([strix.qwen36.interactive2.c8.completion-nothink](https://agmind.ai/claims/strix.qwen36.interactive2.c8.completion-nothink/)).

Structured automation: strict-JSON tasks completed at
**100.0 % of requests** ([strix.qwen36.structured.c1.task-success-nothink](https://agmind.ai/claims/strix.qwen36.structured.c1.task-success-nothink/)).
Long-context retrieval on documents up to 32k tokens:
**100.0 % of requests** ([strix.qwen36.longctx.c1.needle-success](https://agmind.ai/claims/strix.qwen36.longctx.c1.needle-success/)), with the
prefill bill for that document size at
**33965 ms** ([strix.qwen36.longctx.c1.ttft-32k-en](https://agmind.ai/claims/strix.qwen36.longctx.c1.ttft-32k-en/)) — the number
document workflows should be planning around.

The full measured tables live in
[the benchmark repository](https://github.com/botAGI/strix-halo-llm-benchmarks),
and each figure has a permanent page in [the registry](https://agmind.ai/claims/).

## Does it hold up as a serving box

Three continuous hours at concurrency four moved the decode pace by
**1.6 %** ([strix.qwen36.endurance.c4.itl-drift-180m](https://agmind.ai/claims/strix.qwen36.endurance.c4.itl-drift-180m/)), sustained at
**30.0 ms/token** ([strix.qwen36.endurance.c4.itl-median](https://agmind.ai/claims/strix.qwen36.endurance.c4.itl-median/)) — on both units,
with the same shape. The
[endurance report](https://agmind.ai/reports/strix-halo-sustained-load-3-hours/) is the long
version of the answer to "does this thing throttle if I actually use it".

Beyond the frozen tests, the softer evidence: one of the units has been up
for over a week of continuous mixed duty at the time of writing, serving
models while we work on other things. Uptime is not a benchmark, and we are
not presenting it as one — but a box that needs babysitting does not
accumulate quiet weeks.

## What we cannot vouch for

**Two units are not a batch.** Everything above describes our two machines.
Manufacturing variance, revisions and firmware differences across a
production run are exactly the kind of thing two samples cannot measure, and
our [testbed page](https://agmind.ai/testbed/) states this limit plainly.

**We have not tested it as a desktop.** No display-output testing, no
gaming, no thermals under GPU graphics load — these are serving nodes and we
measured serving.

**No comparison against boxes we do not own.** Other vendors ship the same
silicon, and roundups comparing them are easy to write and impossible to
verify. We measured this one because we have it.

**No price advice.** Hardware prices move; the lab does not publish them.

## If you have one

Run the frozen workloads yourself — [the harness is public](https://github.com/botAGI/agmind-bench)
and the [reproduction contract](https://agmind.ai/reproduce/) means a confirmed independent
run raises a claim's evidence level and a failed one gets published too. A
second, third and fourth GTR9 Pro in other people's hands is how "two units
agreed" turns into something stronger than two units agreeing.

---

Machine-readable claim registry: https://agmind.ai/claims.json · llms.txt: https://agmind.ai/llms.txt