Strix Halo for Local LLMs: Ryzen AI Max+ 395, 128 GB Machines, GPU Memory and Speed

Strix Halo is AMD's Ryzen AI Max+ 395: 16 Zen 5 cores and a 40-unit Radeon GPU sharing up to 128 GB of memory at 256 GB/s. For local LLMs, a 128 GB machine holds gpt-oss-120b or a 4-bit-class Qwen3.8-Flash-Next and generates roughly 30 to 50 tokens per second on small or mixture-of-experts models. Dense 27B models run near 8, and long prompts are slow.

“Strix Halo” is the codename AMD’s own specification page gives for the Ryzen AI Max 300 series. The idea is the one behind Apple silicon and NVIDIA’s GB10: put the CPU and a large GPU on one pool of fast laptop memory, so a small computer can load models far beyond a graphics card’s capacity. Its successor, the Ryzen AI Max+ PRO 495, raises the ceiling to 192 GB.

Chip and platform figures below were read from AMD’s specification pages and ROCm documentation on 14 November 2026, and store listings, release notes and model files were opened the same day. Speeds are either credited measurements or estimates from the method behind our tokens-per-second calculator, and we label which. If you are choosing between this platform, a DGX Spark and a Mac Studio, read our three-way desk-side system comparison instead; this guide stays on Strix Halo itself.

What Strix Halo is: Ryzen AI Max+ 395 specs

Two numbers decide what the chip does for language models. The 128 GB ceiling sets which models load. The 256 GB/s of bandwidth sets how fast they answer, because every generated token reads the model’s active weights from memory once. For scale, NVIDIA lists 1,792 GB/s for the RTX 5090, seven times as much, on a quarter of the memory.

The memory is soldered, so the capacity you buy is permanent. Smaller chips in the family have fewer GPU cores: HP’s configurations list the Ryzen AI Max PRO 390 with 12 CPU cores and 32 GPU cores and the PRO 385 with 8 and 32, and Framework sells a 32 GB Desktop with the Ryzen AI Max 385.

Ryzen AI Max+ PRO 495: the 192 GB successor

People search for a “Ryzen AI Max+ 495”. The product AMD announced on 20 May 2026 is the Ryzen AI Max+ PRO 495, with the PRO 490 and PRO 485 beside it, and AMD’s pages list no non-PRO 495.

The new chip is a capacity upgrade: 50 percent more memory, 7 percent more bandwidth and almost the same GPU. A model that fits both runs at nearly the same speed on either. AMD’s claim of 300-billion-parameter models at 4-bit rests on the 160 GB graphics allocation.

How much memory the GPU can use on Windows and Linux

Installed memory and GPU memory are different numbers on this platform, and the rules differ by operating system.

Windows. AMD’s Variable Graphics Memory (VGM) converts part of system RAM into dedicated graphics memory. You set it in AMD Software under Performance, then Tuning, and restart. AMD’s own example: a 128 GB system ships with 512 MB dedicated, and Windows shares about half of RAM with the GPU on top, 64.3 GB addressable in all. With VGM at 96 GB the GPU can address 112 GB, though AMD says the best performance comes from keeping workloads inside the 96 GB block. That block is taken from what the CPU can use. On the PRO 495, AMD puts the dedicated maximum at 160 GB, and its own image-generation tests ran with VGM at 128 GB.

Linux. There is no partition. AMD’s ROCm documentation explains that the driver maps system memory into the GPU’s address space on demand, and that the limit on those mappings defaults to about 50 percent of RAM. AMD recommends keeping the BIOS reservation small, for example 0.5 GB, and raising the shared limit with its helper tool:

pipx install amd-debug-tools
amd-ttm            # shows the current limit; AMD's 128 GB example reports 62.82 GB
amd-ttm --set 100  # usable shared memory in GB; reboot to apply

The same page lists kernel requirements for Ryzen AI Max chips: Ubuntu 24.04 with the 6.17 hardware-enablement kernel or the 6.14 OEM kernel at the patch levels AMD names, or Linux 6.18.4 or later elsewhere. Fedora 43, Ubuntu 26.04 and Arch Linux 2026.02.01 already include the fixes.

Skipping this step is costly. AMD’s ROCm team benchmarked a 128 GB machine with 64 GB open to the GPU, and the 76 GB Qwen3.5-122B-A10B split 61 percent to the GPU and 39 percent to the CPU, at 8.6 tokens per second.

For planning, we follow the VRAM calculator: 80 percent of installed memory on Linux once the limit is raised, and 90 percent of the dedicated block on Windows, as on a graphics card.

Strix Halo mini PCs, desktops and laptops in November 2026

Mini PCs from smaller brands, such as GMKtec’s EVO-X2, use the same chip. Their store pages rate-limited our requests today, so we quote no models or prices for them here; our mini PC comparison lists the EVO-X2 and its rivals with store prices read on 15 November.

Budget with care. HWBusters notes that Framework’s 128 GB Desktop launched at $1,999 last year, and Framework’s September post says it expects memory pricing to keep rising for every capacity over the next six months.

Which models fit in 64, 96, 128 and 192 GB

Weights take the file size of the published build, and each token of context adds a fixed amount of cache. The 96 GB row is the Windows case, a 128 GB machine with VGM at 96 GB; the others assume Linux with the limit raised.

Sizes are today’s Hugging Face listings from ggml-org, Unsloth and bartowski. Per-token cache follows each model’s attention layout, as in our guide to models by VRAM tier: about 37 KB per token for gpt-oss-120b, 27 KB for Qwen3.8-Flash-Next, 64 KiB for Qwen3.8-27B and 320 KiB for Llama 3.3 70B.

Three points follow:

• 128 GB is the tier the platform was built for. It holds gpt-oss-120b with several long sessions, or Qwen3.8-Flash-Next in a 4-bit-class build. Our Qwen 3.8 hardware guide covers that model’s licence, and the gpt-oss setup guide the other.
• Framework’s own picks go past our reserve. Its page recommends a 52.5 GB build of Qwen3.5-122B-A10B on 64 GB and a 163 GB 2-bit DeepSeek V4.1 Flash on 192 GB, 82 and 85 percent of installed memory. That works on a headless Linux machine doing nothing else.
• 192 GB does not reach the next class at 4-bit. GLM-5.3-Flash at UD-IQ4_XS is 156.8 GB, over the budget and within 4 GB of AMD’s 160 GB graphics ceiling. Our GLM-5.3 deployment guide sizes the larger builds.

Strix Halo LLM software: llama.cpp, LM Studio, Lemonade, Ollama and vLLM

Lemonade’s README describes it as a community project with optimisations from AMD engineers, and its server speaks OpenAI-, Anthropic- and Ollama-style APIs. Strata, the runner that keeps mixture-of-experts weights in RAM, also publishes Strix Halo guides for Linux and Windows.

On ROCm against Vulkan, llama.cpp’s two benchmark threads are worth reading together. For Llama 2 7B at Q4_0, the Vulkan thread lists a Ryzen AI Max+ machine at 54.4 tokens per second, reading prompts at 1,704; the ROCm thread lists one at 49.9 and 1,004. Those are different users and builds, so run both yourself. The Vulkan thread asks Linux users to set RADV_PERFTEST=nogttspill:

RADV_PERFTEST=nogttspill llama-bench -m gpt-oss-120b-MXFP4.gguf -ngl 99 -fa 1 -p 4096 -n 256

Our comparison of vLLM, Ollama and llama.cpp explains when each engine is the right one.

Strix Halo LLM speed: measurements and estimates

Our estimates divide 60 percent of peak bandwidth by the bytes each token reads, for one user with a 2,048-token prompt and a 512-token reply:

The measurements sit above the estimates for small dense models: AMD’s Qwen3.5-9B figure, with a 6.2 GB file, works out at about 72 percent of peak bandwidth. Nobody we found has published a single-user figure for gpt-oss-120b or Qwen3.8-Flash-Next on this chip under stated settings, so treat those rows as estimates and measure your own.

Limits: bandwidth-bound generation and slow prompt processing

• Dense models are slow. A dense 27B model reads about 19 GB per token, which 256 GB/s turns into single-digit speeds. Mixture-of-experts models read only their active experts, which is why the 35B model with 3B active parameters ran at 42 tokens per second and why gpt-oss-120b is the natural large model here.
• Prompt processing is the larger gap. Reading a prompt is compute-bound, and AMD rates the GPU at 60 FP16 TFLOPS. On Llama 2 7B, llama.cpp’s threads list 1,704 prompt tokens per second for this chip under Vulkan and 14,970 for an RTX 5090 under CUDA. AMD’s own cost model assumes 446 input tokens per second at 128K context, which makes a cold 128K-token prompt a wait of nearly five minutes.
• Independent testers agree. Tom’s Hardware ran the dense Qwen 3.8 27B on a Ryzen AI Halo in September and called it the worst case in its line-up, with low bandwidth and low prompt speed, adding that multi-token prediction lifted generation but could not offset the prompt times.
• Drafting helps generation only. Framework measured 16 to 20 percent more output with DSpark speculative decoding, and slightly slower prompt reading.
• Concurrency adds throughput, not speed. StorageReview’s vLLM run went from 25 tokens per second for one request to 407 across 64, shared between them.
• The stack has requirements. Kernel versions, the memory limit and the choice of ROCm or Vulkan all change results, more so than on a CUDA machine.

Who should buy a Strix Halo machine for local LLMs

• Buy 128 GB if you want 100B-class mixture-of-experts models on a quiet x86 machine that runs ordinary Linux or Windows software, and you mostly chat, evaluate models or run a few background agents.
• Buy 64 GB for 27B to 35B models at 8-bit with full context. A 32 GB graphics card runs the 4-bit versions several times faster, as our RTX 5090 guide shows.
• Buy 192 GB only if a specific model needs it: a 4-bit Qwen3.8-Flash-Next with many sessions, or 2-bit builds of the next class. The Framework kit costs about twice the 128 GB price HWBusters quotes for 50 percent more memory and about the same speed.
• Look elsewhere for prompt-heavy coding agents, fast dense models or a shared team service. Our guide to graphics cards for local models covers discrete GPUs.

How VDF AI fits

VDF AI does not sell, resell or host Strix Halo machines and has no AMD-specific integration. Its self-hosting requirements size the platform’s own services by vCPU, memory and storage, with Linux as the default production target.

A Strix Halo box is one person’s model host. A team or a regulated organisation also has to decide who hosts the model, who approves models and tools, who may call them, what is recorded and where data stays. VDF AI Router can register an Ollama endpoint or another custom on-premises deployment beside the cloud models policy permits, with air-gap mode restricting routing to local models, while the platform adds governed agents, role-based access control and audit logs. To discuss a pilot, contact us.