Strix Halo AI Benchmarks
Real-world local LLM performance numbers from our Strix Halo dev machine — an AMD APU with 96 GB of unified memory running open-source models via Vulkan. Updated as we test new models.
Hardware
Chip
AMD Ryzen AI Max+ 395
Memory
96 GB LPDDR5X
iGPU
Radeon 890M (40 CUs)
Compute
~16 TFLOPS FP16
Backend
Vulkan (ROCm unavailable on this hardware)
Software
llama.cpp + ollama (Vulkan build)
Model Performance
Benchmarks incoming
We're running tests — results will appear here as models are evaluated.
Gen t/s = generation tokens per second (higher = faster responses). PP t/s = prompt processing speed. Tests run on a single machine with no other heavy GPU load.
Task-Effectiveness Evaluations
Beyond raw tokens/sec: which model actually gets a given job done, at what cost, against what we already use in production. These are our own evaluations — methodology (hardware, run count) is stated with each result, no synthetic benchmark theater.
Compared to: qwen3-coder:30b, Sonnet 5
qwen3.8:27b ~30 tok/s vs qwen3-coder:30b ~20 tok/s on GB10; Sonnet 5 remains the quality ceiling for agent-lane tasks, qwen3.8 selected as the free execution lane for batch/extraction/monitor work under a 3-green verdict verification rule.
qwen3.8:27b ~30 t/s vs qwen3-coder:30b ~20 t/s on GB10; wired as `local-coder-xl` in litellm-config.yaml, routing proven via Ollama POST delta. Sonnet 5 remains the quality ceiling; qwen3.8 selected as the free execution lane for tool-calling/structured-output agent work. Cost: $0 (local) vs roughly $1-3 notional per 1M-token agent run on Sonnet 5. Routing decision: local-coder-xl is the first candidate for batch/extraction/monitor work (fleet rubric section 2, Qwen default); every verdict from it is still checked under the per-agent 3-green verdict rule before being trusted.
GB10 (Grace Blackwell, 128GB) + strix-halo pool, Ollama · 2026-08-21