Guide
OpenClaw on AMD Strix Halo: 128GB Changes What Fits
The AMD Ryzen AI Max+ 395 pairs 16 Zen 5 cores with a 40-compute-unit Radeon 8060S and up to 128GB of LPDDR5X that the CPU and GPU share. For local agents the headline is not speed — it is that models which simply will not load on a 24GB card fit here with room to spare. Here is what that actually buys, and what it costs you.
The One Thing It Does That a 4090 Cannot
An RTX 4090 is faster per token. It also has 24GB of VRAM, and a 70B model at 8-bit needs roughly 57GB of weights before you have loaded a single token of context. That model does not run on a 4090 at that precision — not slowly, not at all.
On a 128GB Strix Halo the same model loads with about 58GB left over for context. The comparison is not “which is faster” but “which one can open the file.” For agent work, where a long session accumulates a lot of context, that headroom is the whole argument.
Real Numbers
| Workload | Throughput |
|---|---|
| 30B mixture-of-experts models | ~100 tokens/sec |
| Qwen3 30B MoE | 66–72 tokens/sec |
| LFM2-24B-A2B | ~109 tokens/sec |
| 70B dense at 8-bit | Loads with ~58GB free for context |
Read the pattern rather than the individual figures: sparse MoE models fly here because only a fraction of their parameters activate per token, while large dense models load comfortably but generate more slowly. Unified memory gives you capacity; it does not give you a workstation GPU's bandwidth.
It Runs Muse Glimmer at Full Precision
A concrete case worth calling out. Meta's Muse Glimmer 30B, the open agent model released in August 2026, needs 64GB for the full BF16 build — which no consumer graphics card offers. A 128GB Strix Halo runs it unquantized with the rest of the memory free for a 128K context.
That combination is close to the practical ceiling for a self-contained local agent right now: an Apache 2.0 model tuned for tool use, running at full precision, on a box on your desk, with nothing leaving the building.
Allocating the Memory
The 128GB is not automatically available to the GPU — you assign it, and this is the step people miss before concluding the machine cannot load their model.
- Windows: allocate up to 96GB to the iGPU through AMD Adrenalin.
- Linux: set a GTT kernel parameter and keep the BIOS framebuffer small, letting the driver hand out the rest dynamically.
Then run the model through LM Studio, Ollama or llama.cpp on the Vulkan backend, which is the path with the fewest surprises on this hardware today. Setup for each is in OpenClaw + Ollama, LM Studio, and llama.cpp.
Where It Sits Against the Alternatives
| Option | Strength | Weakness |
|---|---|---|
| Strix Halo 128GB | Capacity; runs 70B-class models | Memory bandwidth caps dense-model speed |
| RTX 4090 / 5090 | Fastest per token under 24–32GB | Cannot load large models at all |
| NVIDIA DGX Spark | CUDA ecosystem, unified memory | Costs more |
| Hosted API | No hardware, no idle power | Per-token cost, data leaves your network |
The Honest Trade-Off
A local box makes sense when the reason is privacy, an unmetered token budget, or simply wanting to own the thing. It is a worse deal if the goal is a bot that answers messages reliably at 3am: that means a machine that never sleeps, survives your power cuts and router reboots, and gets patched by you.
localhost resolves to the container, not to your house. Pairing the two means exposing your model server on a public HTTPS address, which is a real security decision. The clean options are: run the agent locally too, or connect it over a private network such as Tailscale.Plenty of people land on both — a hosted agent for always-on chat, and the local box for heavy or sensitive jobs. Nothing stops you running the two side by side.
FAQ
Can the Ryzen AI Max+ 395 run a 70B model?
Yes. At 8-bit a 70B model needs about 57GB of weights, which fits in the 128GB unified memory with roughly 58GB left for context. Expect capacity rather than speed — dense models of that size generate slowly here.
Is Strix Halo better than an RTX 4090 for local AI?
It depends entirely on model size. Under 24GB the 4090 is faster. Above it the 4090 cannot load the model at all, and Strix Halo wins by default. For 30B-class agent models the Strix Halo is the more flexible choice.
How much memory should I give the GPU?
Up to 96GB via AMD Adrenalin on Windows. On Linux, set a GTT kernel parameter with a small BIOS framebuffer and let the driver allocate dynamically. Leave enough for the operating system and whatever else the machine runs.
Does OpenClaw run on the Strix Halo box itself?
It can — OpenClaw is light next to the model, so the same machine can host both the agent and the inference server. That is the simplest arrangement and avoids the networking question entirely.