Qwen3.8‑27B + R9700 + 13yo Pentium = ?

It’s time to go big or go home! I’ve bought ASUS Turbo Radeon AI Pro R9700, an RDNA4 graphics card with a whopping 32 GB of GDDR6 VRAM, right before yet another price increase.

Sadly, the only desktop computer I have is an ancient custom build with a 4th-gen Intel Pentium G3420, 8 GB DDR3 RAM, and a single PCIe 3.0 slot. The CPU socket is damaged; the PC works only when I push down on the CPU cooler while booting (so that POST won’t fail). I don’t have any SATA disks anymore, and of course NVMe wasn’t a thing back in the Haswell days.

Yesterday, a new Qwen3.8 model has been released. It’s a dense model with 27 billion active parameters, vision support, and a built-in MTP (multi-token prediction) layer. If benchmarks are to be believed, it’s better than Opus 4.6. Let’s try it out.

First impressions

I’ve installed Ubuntu 26.04 on an SSD and downloaded all the updates. This cursed setup works just fine, even if some apps report that there may be compatibility issues:

Warning: radv is not a  conformant Vulkan implementation

By the way, a USB‑C cable matters A LOT. I have two such cables, both with no logos or markings. The first one is too slow, at only 42 MB/s. The other one is 10x faster, but every time you look at it wrong, it would lose a connection. Choose your poison, I guess.

I had to update VBIOS to make the graphics card quiet. This update changes the minimum fan speed from 30% to 12% (about 1000 RPM). Unfortunately, VBIOS can be updated only on Windows, but you can temporarily install Microsoft malware directly on SSD without losing any data.

There is no official AMD control panel for Linux, but you can use a wonderful LACT app instead. I’ve set the fan curves to start at 12% and never exceed 40%. You can also change power limits. The default is 300 W, but I haven’t noticed any performance issues at 210 W, and it helps keep the temperatures below 85°C without thermal throttling.

Testing

Qwen3.8 has multiple reasoning modes, with xhigh being the default. It behaves like a caveman, which is a downgrade for me. With Qwen3.6 I could analyze the reasoning and refine the prompt early if the model hallucinated.

We need answer user question. Need analyze code, explain functionality, rate quality. We need think in English same language. Need inspect code carefully. (…)

How long would it take to process a large chunk of code and answer these two questions:

• What does this code do?
• How would you rate it?

Flathub’s moderation code is approximately 2350 lines of text (including empty lines) in a 90 kB file, which is about 19700 input tokens. I don’t think the model has been benchmaxxed against this file, so it should show the real performance. I’ve limited the output to 8192 tokens and applied other llama-cli parameters:

--no-mmap
--n-gpu-layers all
--flash-attn on
--fit off

--ctx-size 131072
--cache-type-k q8_0
--cache-type-v q8_0
--temperature 1.0
--top-k 20
--top-p 0.95
--min-p 0
--repeat-penalty 1
--presence-penalty 0
--reasoning on
--reasoning-effort xhigh

--spec-type draft-mtp
--spec-draft-type-k q8_0
--spec-draft-type-v q8_0
--spec-draft-n-max 1
--n-gpu-layers-draft all

Tested with the official llama.cpp Vulkan build b10434, AMD Lemonade llamacpp-rocm build b1311, and Unsloth GGUF quants, with 128k context size and average MTP acceptance rate of 0.65 at 1 token ahead (2 is barely okay). Sampler parameters are based on the official recommendations. Note that Q8_K_XL quant won’t fit in 32 GB VRAM.

Results

Vulkan

(This table can be scrolled horizontally on mobile.)

ROCm

(This table can be scrolled horizontally on mobile.)

*) With --no-mmap removed due to OS stability issues

**) With 96k context size, 128k won’t fit

Observations

Qwen3.8 is based on a battle-tested qwen35 architecture and requires no custom patches for llama.cpp. It doesn’t crash, it has valid reasoning (with primitive language), and responses are correct and coherent. MTP works, resulting in huge performance improvements on Vulkan. ROCm has faster prompt processing, but token generation speed is similar, and it requires more VRAM.

The Q6_K_XL quant is slower than the 8-bit version, which is interesting.

People report that they can generate up to 30 tokens per second as well. So it looks like you can buy a powerful graphics card, pair it with mediocre components, and still be happy. No need to pay extra for expensive DDR5 RAM, just spend it on a better GPU.