Day-zero support for the new Qwen3.8 27B just landed on two AMD boxes — one needs a graphics card, one doesn't — and the tokens-per-second numbers are real.
On August 14, AMD flipped on day-zero support for Qwen3.8 27B — Alibaba's newest dense model — across two very different pieces of hardware: a $1,299 Radeon AI PRO R9700 workstation card, and, more interestingly, a Ryzen AI Max+ 395 processor with no discrete GPU attached at all. AMD's own numbers: 51.8 tokens per second on the R9700, 24.5 tokens per second on the Ryzen chip's integrated graphics alone. Both are fast enough to actually use, not just fast enough to demo.
What Qwen3.8 27B actually is
Qwen3.8 27B is a dense 27-billion-parameter model — not a mixture-of-experts setup where only a slice of weights activates per token — built around a hybrid of Gated DeltaNet and Gated Attention layers aimed at longer context handling. It hit Hugging Face the same day AMD shipped support for it, which is the actual news here: this isn't a driver update landing six months after a model's release, it's the same day. Dense 27B is a specific, deliberate size — big enough to be genuinely useful, small enough to fit inside 24GB, which happens to be exactly what a lot of prosumer and workstation hardware already ships with.
27B params
Model size
dense, not MoE
~24GB
Memory needed
to load the model
51.8 tok/s
R9700 speed
Vulkan backend
24.5 tok/s
Ryzen AI Max+ speed
integrated graphics only
Two boxes, two very different bets
The R9700 route means buying a $1,299 workstation card and dropping it into a PC you already own — 32GB of dedicated VRAM, normal PCIe bandwidth, a normal desktop wrapped around it. The Ryzen AI Max+ 395 route means buying an entire small-form-factor PC built around the chip; Framework's Desktop with 128GB of unified memory currently runs $3,449, up from a $1,999 launch price as RAM costs climbed through 2026. One is a component upgrade. The other is a whole-machine purchase — quieter, cooler, no separate GPU or PSU headroom needed. Compare either against the $16,000 RTX PRO 6000 towers built for 200B-class models, and for a 27B dense model, that kind of spend is obvious overkill.
R9700 card vs Ryzen AI Max mini PC
Price
$1,299 (card only)
Radeon AI PRO R9700
~$3,449 (full 128GB system)
Memory
32GB dedicated GDDR6
Radeon AI PRO R9700
Up to 128GB unified LPDDR5X
Qwen3.8 27B speed
51.8 tok/s
Radeon AI PRO R9700
24.5 tok/s
What you need
An existing PC with a PCIe slot and PSU headroom
Radeon AI PRO R9700
Nothing else — it's the whole computer
Best for
Existing desktop owners wanting dedicated VRAM
Radeon AI PRO R9700
Buyers who want one small, quiet box
Radeon AI PRO R9700
Ryzen AI Max+ 395 mini PC
Price
$1,299 (card only)
~$3,449 (full 128GB system)
Memory
32GB dedicated GDDR6
Up to 128GB unified LPDDR5X
Qwen3.8 27B speed
51.8 tok/s
24.5 tok/s
What you need
An existing PC with a PCIe slot and PSU headroom
Nothing else — it's the whole computer
Best for
Existing desktop owners wanting dedicated VRAM
Buyers who want one small, quiet box
AMD's Radeon AI PRO R9700 ships with 32GB of dedicated memory for about $1,299. · Unsplash
The part the press release won't tell you
None of this runs on CUDA. It runs through Vulkan and AMD's ROCm stack, via LM Studio or the open-source llama.cpp project — both genuinely good, both improving fast, but still a step behind the plug-and-play state of Nvidia's CUDA ecosystem for anything outside the mainstream model families. I switched from Nvidia to AMD for a different card earlier this year, and the software gap was the real surprise, not the hardware. Day-zero support for a brand-new model is a genuinely good sign AMD is closing that gap — but it's closing it, not closed.
Install LM Studio (free) or set up AMD's Lemonade server
2
Pull the Qwen3.8 27B GGUF build from Hugging Face
3
Select the Vulkan backend in the runtime settings
4
Load the model — expect roughly 24GB of memory use before context
Verdict
Buy call
The R9700 wins on price-per-token if you already have a PC. The Ryzen AI Max mini PC wins if you want a single silent appliance that might scale to bigger models later.
Best for: Existing desktop owners → R9700. Everyone else wanting one clean box → Ryzen AI Max mini PC.
Can I run Qwen3.8 27B without a discrete GPU?
Yes — AMD's Ryzen AI Max+ 395 runs it on integrated graphics alone at 24.5 tokens per second, per AMD's own day-zero benchmarks.
How much does it cost to get started?
The cheapest path is a $1,299 Radeon AI PRO R9700 card in a PC you already own. A full Ryzen AI Max mini PC with 128GB starts closer to $3,449 as of August 2026.
Does this work on Nvidia hardware?
Qwen3.8 27B isn't AMD-exclusive — it runs anywhere with roughly 24GB of memory, including a 16GB-class Nvidia card at lower quantization. AMD's day-zero support is about tuned performance, not exclusivity.
Is AMD's software actually ready for this?
Mostly. LM Studio and Vulkan handle mainstream models like Qwen3.8 well, but ROCm's coverage of edge-case models and tooling still trails CUDA. Budget for some troubleshooting if you go off the beaten path.
Day-zero support for a 27B model on a $1,299 card is a genuinely different AMD than the one that used to ship drivers a year late. Whether that's enough to actually pull buyers away from Nvidia is still an open question — earlier Meta Muse Glimmer 30B benchmarks showed the gap narrowing but not closed. If you're building a local-AI box for under $1,500 total, the R9700 just became one of the more interesting options on the table.
