More
Best rig by budgetRent before you buySubmit a buildNewsletterFree datasetAdvertise
Submit build

Builds / AMD Radeon 8060S (Strix Halo, unified)

GMKtec EVO-X3 (Strix Halo 128 GB, Windows 11, stock llama.cpp Vulkan): Qwen3.6 35B-A3B at 61.8 tok/s, Qwen3.5 122B at 23.6 (idle machine)

Last updated:

Source: github.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
GPU(s)
AMD Radeon 8060S (Strix Halo, unified)
VRAM / unified memory (GB)
128
CPU
AMD Ryzen AI Max+ 395 (16 cores / 32 threads)
Amazon · eBay (used) · Newegg
Motherboard
GMKtec EVO-X3 mini PC
Amazon · eBay (used) · Newegg
RAM
128 GB at 8000 MT/s (16 GB reserved for graphics; Windows sees 111.6 GB)
Amazon · eBay (used) · Newegg
PSU
not stated
Other
Windows 11 Pro 10.0.26300; stock llama.cpp b11401 (commit a7fb71f) prebuilt Windows x64 Vulkan zip, unmodified; no ROCm, no Linux; Vulkan 1.4.329, device-local heap 70.21 GiB (budget 66.70 GiB); -ngl 99, one slot, -ctxcp 0 -cram 0; flash attention and micro-batch left at defaults
Reported total cost
not stated
Power
not stated

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Qwen3.6 35B-A3BUnsloth UD-Q4_K_XL (22.4 GB)llama.cpp b11401 (Windows x64 Vulkan, unmodified)61.8not statedcontext: short chat request, thinking off; 32K context configured for the first-token sweep📏 Median of 30 warm requests to a loaded llama-server on loopback after one cold request; thinking off; Machine B ran nothing else. Machine A (also running a voice assistant) was 55.3 tok/s. Updating Machine B's driver to 32.0.32015.2008 left speed unchanged at 60.6 tok/s. Warm first-token median 145 ms (p95 2,311 ms) on Machine B. No speculative decoding.
Qwen3.5 122B legacyUnsloth UD-IQ4_XS (60.2 GB / 56.1 GiB, 3 shards)llama.cpp b11401 (Windows x64 Vulkan, unmodified)23.6not statedcontext: short chat request, thinking off; 32K context configured for the first-token sweep📏 Median of 30 warm requests; Machine B idle 23.6 tok/s, Machine A 23.1 tok/s. Warm first-token median 425 ms on Machine B. IQ4_XS was picked because the usual 4-bit files (UD-Q4_K_XL 71.74 GiB, Q4_K_M 71.28 GiB) exceed the 66.70 GiB budget Vulkan reports; those larger files were not tested. No speculative decoding.

llama.cpp GitHub discussion #31839 by IanCrossland, posted 7 Nov 2026 (measured in October 2026). Two identical GMKtec EVO-X3 mini PCs; headline figures are the idle Machine B medians. Both models fit in memory at once with 27 GiB of Windows memory to spare; generating on both at once halved the 35B (47.9 alone → 24.0). A ~19K-token prompt waited about 19 s on the 35B (with -cram 0) and 55 s on the 122B for its first token. Author: every other Strix Halo LLM benchmark they found was run on Linux.

Source: https://github.com/ggml-org/llama.cpp/discussions/31839

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.

Markdown

[![Measured on RigBase](https://getrigbase.com/badge/gmktec-evo-x3-strix-halo-128gb-windows-vulkan-31839.svg)](https://getrigbase.com/builds/gmktec-evo-x3-strix-halo-128gb-windows-vulkan-31839/)

HTML

<a href="https://getrigbase.com/builds/gmktec-evo-x3-strix-halo-128gb-windows-vulkan-31839/" rel="noopener"><img src="https://getrigbase.com/badge/gmktec-evo-x3-strix-halo-128gb-windows-vulkan-31839.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase · How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Quick answers: Can Strix Halo (Radeon 8060S) run Qwen3.5 122B? · Can Strix Halo (Radeon 8060S) run Qwen3.6 35B-A3B?