Strix Halo Guide

Strix Halo Guide is a comprehensive, practical guide to running large language models locally on AMD Strix Halo / Ryzen AI MAX+ 395 systems with Radeon 8060S (gfx1151) and 96GB/128GB unified memory. Maintained by community member hogeheer499-commits, it covers BIOS configuration, Ubuntu 24.04/kernel setup, Ollama, llama.cpp Vulkan/RADV, ROCm/HIP experiments, vLLM, benchmarks, raw logs, and reproducibility checks¹².

With 295 stars, it is one of the most popular community resources for Framework Desktop local AI workloads. The guide is actively maintained and covers Framework Desktop, Beelink GTR9 Pro, GMKtec, and other Strix Halo systems¹.

What It Covers

The guide provides copyable setup commands, measured benchmark rows, raw evidence, and community reproductions for running local LLMs on Strix Halo hardware¹:

• BIOS config — UMA Frame Buffer settings, IOMMU policy
• Ubuntu 24.04 / kernel setup — GRUB kernel parameters (amdgpu.gttsize, ttm.pages_limit), tuned accelerator-performance profile, Mesa via kisak PPA
• Ollama — Vulkan backend configuration (OLLAMA_VULKAN=1, HIP_VISIBLE_DEVICES=-1)
• llama.cpp — Vulkan/RADV as the primary path, ROCm/HIP experiments
• ROCm — HIP integrated-device support, ROCm_Host model buffers on gfx1151
• vLLM — Notes and compatibility alerts
• Benchmarks — Raw CSVs, logs, charts, and reproducibility notes

Setup Script

An automated setup.sh script installs the Linux-side Vulkan/RADV + Ollama path, configures Ollama for Vulkan, pulls a model, and prepares a verification benchmark¹:

git clone https://github.com/hogeheer499-commits/strix-halo-guide
cd strix-halo-guide
bash setup.sh

Headline Benchmarks

Model	Backend	Result	Notes
Qwen3-Coder 30B-A3B Q4_K_S	llama.cpp Vulkan/RADV (b9851)	100.99 t/s	Direct llama-bench, not MTP/server
Qwen3-30B-A3B-Instruct-2507 IQ4_XS	llama.cpp Vulkan/RADV (b9467)	100.04 t/s	b9544 control at 103.18 t/s
LFM2.5 8B-A1B Q4_K_M	llama.cpp Vulkan/RADV	170.02 t/s	Small active-parameter MoE speed result
DeepSeek V4 Flash 284B UD-IQ2_XXS	llama.cpp Vulkan/RADV (b10034)	13.27 t/s	90.86GB capacity scout, loaded directly
Nemotron 3 Super 120B-A12B UD-IQ4_XS	llama.cpp Vulkan/RADV	18.43 t/s	120B-class capacity route
Qwen3.6 35B-A3B Q4_K_M	Ollama 0.31.2 (Vulkan)	60.57 t/s	Warm API generation, reboot-qualified
Qwen3.8 27B Q4	Ollama (Vulkan)	—	Qualified with dedicated buyer route and freshness release; Ollama iGPU env documented in quick start1

All headline claims are backed by CSVs, raw logs, charts, and reproducibility notes in the repository's data/ directory¹.

Community Validation

The evidence map covers 11 systems or independent sources from 8 credited community benchmark contributors, with first-party and community claims kept separate. Contributors include Beelink, Corsair, GMKtec, MS-S1-Max, Nimo, NixOS/NPU, and ROCmFP4 system owners. A second independent Beelink GTR9 Pro reproduction confirmed the -ub clamp finding¹.

Upstream Contributions

The maintainer has accepted upstream contributions to several AI infrastructure projects¹:

• llama.cpp (PR #25643)
• LocalAI
• Qwen Code
• OpenTelemetry GenAI
• NVIDIA AICR
• vLLM GGUF plugin (tested-coverage documentation)

Coverage Areas

Topic	Documentation
Thermal stability	THERMAL_STABILITY.md — Corsair three-system evidence, SCLK tradeoffs, fan-reset patch
MoE concurrency	MOE_CONCURRENCY.md — Concurrency-8-to-9 cliff, AMD/RADV density gate
USB4 cluster tuning	USB4_CLUSTER_TUNING.md — Multi-node RPC, cluster latency
Reproducibility	REPRODUCIBILITY.md, data/headline_claims.csv — Dated claims with CSV, raw logs, charts
Best known profiles	BEST_KNOWN_PROFILES.md, data/best_known_profiles.csv — Machine-readable workload-to-runtime recommendations
Fine-tuning	UNSLOTH_STRIX_HALO.md — ROCm 7.2 Unsloth container, one-step SFT smoke, GGUF export
ROCm/vLLM bugwatch	ROCM_VLLM_BUGWATCH.md — Current Strix Halo compatibility alerts
Mini PC comparison	Strix Halo mini PC buying comparison page — cross-vendor evaluation
CPU fallback	Quick start CPU-fallback troubleshooting for Ollama iGPU env

Technical Details

Detail	Value
License	MIT
Stars	295
Forks	20
Commits	267+
Website	strixhaloguide.com
Primary test hardware	Beelink GTR9 Pro (Ryzen AI MAX+ 395, Radeon 8060S, 128GB LPDDR5X-8000)
Community contributors	8 benchmark contributors across 11 systems/sources
Setup script	setup.sh (automated Ubuntu + Vulkan/RADV + Ollama)
Last updated	August 25, 2026 (actively maintained)

Related Projects

Project	Description
framework-rocm-containers	Reproducible ROCm containers for Framework Desktop (PyTorch/JAX)
strix-llm	Known-good local LLM inference configuration for Framework Desktop

Footnotes

1. hogeheer499-commits/strix-halo-guide — GitHub ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

2. Using a Framework Desktop for local AI — Framework Community (official blog post) ↩