Setting Up Claude Code on Strix Halo or a Mac Studio M5 Max

The hardware benchmarks piece covers where the 128GB tier sits relative to everything smaller. This is the setup guide for it: two very different machines that land in the same memory class, with different rules for what you can actually run.

Strix Halo (AMD Ryzen AI Max+ 395) on Linux

Buy the 128GB configuration, not the 64GB one, if you want this tier at all. The 64GB EVO-X2’s factory-default GPU carve-out is 32GB - enough for Llama 3.3 70B at Q4_K_M, not much past it.

The OS choice isn’t cosmetic here. On Linux, the GPU can address roughly 108GB of the 128GB pool through a shared GTT allocation. On Windows, the GPU is limited to a fixed BIOS memory carve-out with no equivalent shared pool. If the reason you bought 128GB was 120B-class models or long context on a 70B model, that capability is a Linux-only feature today. Set the BIOS-level GTT/UMA allocation before you install anything - most boards default to something conservative, and you’ll hit an out-of-memory error on model load if you don’t raise it.

From there, the setup mirrors the standard 2026 path:

ollama pull qwen3-coder:30b
export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
claude

Vulkan and ROCm backends both work on Strix Halo; ROCm is the better-supported path for llama.cpp’s MoE offload flags if you’re running a mixture-of-experts model rather than a dense one.

Mac Studio M5 Max

No OS decision to make - macOS’s unified memory model already gives the GPU access to the full pool. The relevant number is bandwidth, not capacity: the higher-core-count M5 Max configuration reaches 614 GB/s, up from 546 GB/s on M4 Max. That’s the more relevant figure for Claude Code specifically, since the agentic loop re-sends accumulated context on every turn - bandwidth-bound work, not a single large batch.

Setup is the same as any Apple Silicon box on the current guide:

brew install ollama
ollama pull qwen3-coder:30b
export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
claude

oMLX is worth installing alongside Ollama here - its two-tier cache is the one backend that avoids the cold-cache penalty on long sessions, which matters more at 128GB where you’re more likely to be running a model large enough that a full re-read is expensive.

Which One to Buy

If you’re on Linux already, or don’t mind being on Linux, Strix Halo at 128GB is meaningfully cheaper - $2,199 to $3,649 depending on vendor, against Mac Studio M5 Max pricing that starts well above that for a comparable memory configuration. If you need Windows, or you want the bandwidth and don’t want to manage GTT allocation and backend selection yourself, the Mac Studio is the simpler machine to own.

Neither one changes what to run once it’s set up. Follow the model recommendations in the 2026 setup guide for your specific parameter count and context needs - this just tells you how to get the memory tier itself working first.

---

Related Reading

• Claude Code Local LLM Speed by Hardware Tier
• Running Claude Code with a Local LLM in 2026: No Proxy Required