What is the best model to run on a Ryzen AI Max+ 395 with 128GB?

Qwen3-30B-A3B (and the Qwen3-Coder variant) is the daily driver — community llama.cpp benchmarks on the Vulkan RADV backend report about 100 tok/s and 98.5 tok/s respectively at 4-bit. If you want maximum capability instead of maximum speed, gpt-oss 120B in MXFP4 fits in the 96-110GB GPU allocation and reports between 31 and 55.6 tok/s depending on backend and build. Both are sparse mixture-of-experts models, which is exactly what this bandwidth-limited box wants.

Can the Ryzen AI Max+ 395 run a 70B dense model?

It fits, but it is slow. Llama 3.3 70B at Q4 reads roughly 40GB of weights per token, and at ~215 GB/s of real memory bandwidth that caps decode near 5 tok/s. The model loads and works, so it is fine for batch or overnight jobs, but it is not usable for interactive chat or agent loops. Use a MoE model with a small active parameter count instead — Llama 4 Scout (109B total, 17B active) runs about 18.3 tok/s on the same hardware for that reason.

How much of the 128GB can the GPU actually use?

On Windows, AMD Adrenalin caps the iGPU allocation at 96GB. On Linux you can push it to roughly 110GB of usable graphics memory via kernel parameters, which is why the Linux path is the one serious users take. Linux also runs faster: the Windows stack performs roughly 20-30% below the Linux Mesa RADV Vulkan path in community testing.

Is a Strix Halo box better value than a DGX Spark or Mac Studio?

On price per gigabyte, yes, and it is not close — a 128GB Framework Desktop is $1,999 and a GMKtec EVO-X2 has sold as low as $1,499, versus roughly $4,000 for a DGX Spark or a 128GB Mac Studio M4 Max. On bandwidth it loses: ~215 GB/s real versus ~273 GB/s on the Spark and 546 GB/s on the M4 Max, so dense models decode faster on both rivals. The Spark also wins on CUDA tooling and can cluster over 200Gb networking; the Mac wins on quiet, working-out-of-the-box software. Buy Strix Halo when you want the most memory per dollar and you will run MoE models on Linux.

← All guides

Hardware August 5, 2026

Best Models to Run on AMD Ryzen AI Max+ 395 Boxes (August 2026)

The Ryzen AI Max+ 395 (Strix Halo) put 128GB of unified memory in a 2-liter box for about $2,000 — roughly half a DGX Spark and a third of a 128GB Mac Studio. The catch is bandwidth: 256 GB/s theoretical, ~210-220 GB/s real. That number decides everything. Sparse MoE models fly (Qwen3-30B-A3B ~100 tok/s, gpt-oss 120B in the 31-55 tok/s range), dense 70B crawls at ~5 tok/s, and prefill on long prompts is the weak spot nobody puts in the marketing. Here is what to actually run, and where the AMD software stack still hurts.

Building a Strix Halo OpenClaw rig?

See our AI training options. We'll plan a Ryzen AI Max+ OpenClaw setup that turns your mini-PC into a private AI server for your team.

Bottom Line (August 2026)

Daily driver — Qwen3-30B-A3B or Qwen3-Coder 30B-A3B at 4-bit: ~100 tok/s and ~98.5 tok/s on the Vulkan RADV backend. This is the model that makes the box feel fast.
Maximum capability — gpt-oss 120B MXFP4, reported between 31 and 55.6 tok/s depending on backend and build. It fits inside the GPU allocation and is the best reason to buy 128GB.
Long context — Llama 4 Scout (109B total / 17B active, Q4_K_M, 60.9GB) at ~18.3 tok/s, and AMD documents it running at a 256K context window with Flash Attention on and a Q8 KV cache.
Do not bother — dense 70B. Llama 3.3 70B Q4 lands near 5 tok/s. It fits; it is not interactive.
Run Linux. The Windows stack sits roughly 20-30% below the Linux Mesa RADV Vulkan path, and Linux unlocks ~110GB of GPU memory instead of Windows’ 96GB cap.

The Hardware, Briefly

Strix Halo is one chip with a 256-bit LPDDR5X-8000 bus — four times the width of a normal Ryzen APU. That gives 256 GB/s theoretical bandwidth, measured around 210-220 GB/s under real mixed load. All 128GB is unified, so the iGPU addresses model weights directly with no PCIe copy.

Bandwidth is the whole story. Decode speed is roughly bandwidth divided by bytes read per token. A dense 70B at Q4 reads ~40GB per token, so ~215 GB/s gives you ~5 tok/s and no software fix changes that. A MoE model with 3B active parameters reads a fraction of that, which is how the same box hits 100 tok/s. Buy this machine to run sparse models.

The boxes shipping it, as of August 2026:

Box	128GB price	Note
Framework Desktop	$1,999	Standard-parts chassis; $1,099 entry config is a 32GB Max 385 — skip it
GMKtec EVO-X2	$1,499–$2,199	Cheapest route to 128GB when it is on sale
Beelink GTR9 Pro	$1,899–$1,999	Same silicon, different chassis
Minisforum MS-S1 Max	~$2,299
HP Z2 Mini G1a	Enterprise pricing	Buy this only if you need a support contract
AMD Ryzen AI Halo dev platform	$3,999	AMD’s own developer box

What to Run

Model	Quant	Speed	Why
Qwen3-30B-A3B	IQ4_XS	~100 tok/s	Fastest capable assistant on the box (Vulkan RADV)
Qwen3-Coder 30B-A3B	Q4_K_S	~98.5 tok/s	Same speed, tuned for code — the OpenClaw pick
Qwen 3.6	Q4_0	~81 tok/s	~63 tok/s on the balanced build
gpt-oss 120B	MXFP4	31–55.6 tok/s	Biggest brain that fits; llama-server Q4_K_M reports 53.4 tok/s
Llama 4 Scout (109B/17B)	Q4_K_M, 60.9GB	~18.3 tok/s	256K context with FA + Q8 KV per AMD
Qwen3-235B-A22B	4-bit	~8–11 tok/s	Fits, but 22B active is too many for this bandwidth
7B–13B dense	Q4	30–45 tok/s	Fine, but you did not need 128GB for these
Llama 3.3 70B dense	Q4	~5 tok/s	Bandwidth wall. Batch jobs only

Our pick: Qwen3-Coder 30B-A3B for agent work, gpt-oss 120B when quality matters more than latency. The spread between them is the real decision on this hardware. Scout’s 18.3 tok/s hits about 73% of the theoretical ceiling for a 17B-active model at 215 GB/s, which tells you the runtime is already close to optimal — the silicon is the limit, not the software.

Where the Software Still Hurts

This is the part the spec sheet hides.

Prefill is the weak spot. Decode numbers get quoted everywhere; prompt processing does not. Feeding a long prompt takes seconds to minutes at large context. Community testing puts ROCm 7.x nightlies at roughly 502 tok/s prompt processing — best in class for this box, and still an order of magnitude below a discrete GPU. If your OpenClaw workflow re-reads a large codebase every turn, you will feel this before you feel decode speed.

ROCm is preview-grade on this chip. Strix Halo is gfx1151, and getting ROCm running still means setting HSA_OVERRIDE_GFX_VERSION=11.5.1. The pragmatic split right now: Vulkan RADV for token generation, ROCm nightlies for prompt processing. Most people run llama.cpp or Ollama on Vulkan and skip the full ROCm stack entirely.

Windows costs you 20-30% versus Linux RADV, plus 14GB of addressable GPU memory. If you buy this box and run Windows, you bought a slower box.

Agent tooling has sharp edges. Ollama v0.14.0 (January 2026) added Claude Code compatibility, but its OpenAI compatibility layer ignores tool_choice, which can send agents into loops. Test your harness before you trust it with a long run.

Strix Halo vs DGX Spark vs Mac Studio

All three sit at 128GB. They are not the same machine.

	Ryzen AI Max+ 395	DGX Spark	Mac Studio M4 Max
128GB price	$1,499–$2,199	~$4,000	~$3,999
Real bandwidth	~215 GB/s	~273 GB/s	546 GB/s
Ecosystem	Vulkan / ROCm preview	CUDA, NVFP4, vLLM	MLX, works out of the box
Clustering	USB4/Thunderbolt RPC (~9.4 Gbps)	200Gb ConnectX, real TP	None

Buy Strix Halo for the most unified memory per dollar, if you will run MoE models on Linux and can tolerate rough edges. It is roughly half the price of the other two.

Buy the DGX Spark if you want CUDA tooling and a growth path — two Sparks link over 200Gb ConnectX into a real 256GB tensor-parallel machine. Strix Halo’s distributed option is llama.cpp RPC over a ~9.4 Gbps Thunderbolt link, which works but is a different class of interconnect.

Buy the Mac Studio if bandwidth and silence matter more than price. 546 GB/s is 2.5x Strix Halo, so dense models that crawl on AMD are usable on the Mac.

The honest summary: Strix Halo is the value play, not the performance play. At $1,999 for 128GB it is the cheapest way to hold a 120B-class model in memory. You pay for that with prefill speed and a software stack you have to babysit.