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 it still wins, but by much less than it did. In October 2026 the GMKtec EVO-X2 128GB is $3,499.99 and the Minisforum MS-S1 MAX is $3,799 on the vendor stores. The Beelink GTR9 Pro 128GB/2TB is $4,349, and the Framework Desktop 128GB is $3,449 but out of stock. A DGX Spark is $6,950 since 2 October 2026, up from $4,699 (and $3,999 at launch). A Mac Studio M5 Max with 128GB is $5,099 on apple.com. On bandwidth Strix Halo still loses: ~215 GB/s real versus ~273 GB/s on the Spark and 614 GB/s on the 40-core M5 Max, so dense models decode faster on both rivals. Buy Strix Halo when you want the most memory per dollar and you will run MoE models on Linux — just do not expect it to be half the price of the alternatives any more.

← All guides

Hardware August 5, 2026

Best Models to Run on AMD Ryzen AI Max+ 395 Boxes

The Ryzen AI Max+ 395 (Strix Halo) put 128GB of unified memory in a 2-liter box. That box cost about $2,000 at launch. In October 2026 the cheapest in-stock 128GB box is $3,499.99, against $6,950 for a DGX Spark (since 2 October 2026) and $5,099 for a 128GB M5 Max Mac Studio. Strix Halo is still the cheapest 128GB. It costs about half a Spark, but about 69 percent of the Mac Studio. The other 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.

Bottom Line

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 that ship it, with vendor-store prices read on 3 October 2026. Every number here went up during the 2026 DRAM shortage. Treat these as a snapshot, not a price list:

Box	128GB price (Oct 2026)	Note
Framework Desktop	$3,449 (out of stock)	Was $1,999 at its Feb 2025 launch. +73%, almost entirely RAM and storage cost. Pre-order only on 2026-10-03. Standard-parts chassis
GMKtec EVO-X2	$3,499.99 (1TB) / $3,649.99 (2TB)	Vendor store, in stock, 2026-10-03. The cheapest in-stock 128GB box. The August $1,999.99 price was sold out
Beelink GTR9 Pro	$4,349 (2TB)	Vendor store, in stock, 2026-10-03. The 4TB SKU is a pre-order from $4,549. At this price, a $6,950 DGX Spark costs $2,601 more
Minisforum MS-S1 Max	$3,799 (2TB) / $4,119 (4TB)	Vendor store, in stock, 2026-10-03. Launched at $2,299
HP Z2 Mini G1a	Enterprise pricing	Buy this only if you need a support contract
AMD Ryzen AI Halo dev platform	check current listing	AMD’s own developer box; the widely quoted $3,999 predates the 2026 hikes

The Framework line is the one to notice. It was the poster child for cheap 128GB and it has nearly doubled. If you read a 2025 guide recommending it at $1,999, that recommendation was priced against a machine you can no longer buy at that price.

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.

Update, October 2026: AMD’s ROCm 10.1 data now lists the Ryzen AI Max+ 395 (gfx1151) as supported, on Ubuntu 26.04 or 24.04.4 with the HWE 6.17 kernel, so the override is no longer the starting point. The current setup, with both backends and the kernel memory settings, is in AMD ROCm Setup for Local LLMs.

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 (M5 Max)
Price at 128GB (Oct 2026)	$3,499.99–$4,349	$6,950 (was $4,699; $3,999 at launch)	$5,099 (M5 Max, 40-core GPU)
Bandwidth	~215 GB/s real	~273 GB/s	614 GB/s (M5 Max, 40-core GPU, Apple spec)
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 used to be roughly half the price of the other two. In October 2026 the in-stock GMKtec and Minisforum boxes cost 50-55% of a DGX Spark, after NVIDIA’s October raise, and 69-75% of a 128GB Mac Studio. Still the cheapest 128GB, but the price argument no longer carries the decision on its own — if you were choosing Strix Halo purely because it was half price, re-run that comparison.

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. The 40-core M5 Max is rated at 614 GB/s, almost 3x Strix Halo, so dense models that crawl on AMD are usable on the Mac. The 128GB configuration is $5,099. Apple replaced the line on 25 August 2026, and the M4 Max and M3 Ultra are discontinued. See Best Local LLM for Mac Studio for the new tiers.

The honest summary: Strix Halo is still the value play, and the value is thinner than it was. At $3,499.99 for an in-stock 128GB box (October 2026) it remains the cheapest way to hold a 120B-class model in memory, and it is now the only desktop way to hold one in unified memory under $4,500. You pay for that with prefill speed and a software stack you have to babysit.