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

GMKtec EVO-X2 (Strix Halo 128 GB, quiet 54 W mode): Qwen3.8 Flash-Next UD-Q4_K_XL at 37.2 tok/s on Gufo ROCm (MTP), Qwen3.6 35B-A3B at 81.9 with MTP

Last updated: 9 Nov 2026

Source: bryanlurer.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 (16C/32T)
Amazon · eBay (used) · Newegg
Motherboard: GMKtec NucBox EVO-X2 (Sixunited AXB35-02)
Amazon · eBay (used) · Newegg
RAM: 128 GB unified (62 GiB OS-visible + 64 GiB GPU reservation)
PSU: not stated
Other: Fedora 44 KDE, kernel 7.2.8; Vulkan 1.4.354 with Mesa RADV 26.2.3; llama.cpp 0.6.0-dev build 11548 (Vulkan); Gufo v0.31.0 (be7bd3b, ROCm 10.1, rootless podman) for Flash-Next; embedded-controller power mode 'quiet' (STAPM 54 W / fast PPT 100 W / slow 54 W) via the ec_su_axb35 module with custom fan curves
Reported total cost: not stated
Power: quiet mode: STAPM 54 W

Measured speed (generation tokens/sec)

Model	Quant	Engine	tok/s	Prompt processing (tok/s)	Details
Qwen3.8 Flash-Next	Unsloth UD-Q4_K_XL (about 112 GB) + shared Q8_0 MTP head	Gufo v0.31.0 (ROCm 10.1), --speculative mtp --sessions 1	37.2	1,251 (prompt length not stated)	⚡ spec-decode (MTP)context: Gufo model-bench single-mtp, depth 0 ('mixed' text)📏 Gufo's own model-bench --table single-mtp --repetitions 3, measured 2026-11-06: tg mixed 37.2 ± 0.8 at depth 0, 38.9 at 8K, 38.8 at 32K, 41.0 at 131K; repetitive text 132.9 at depth 0 (not used as the headline). Streaming check through the OpenAI API: 36-42 tok/s on prose, 60-69 on code. Loads in 40 s, uses 92 GB of GPU memory
Qwen3.6 35B-A3B	Unsloth UD-Q4_K_M with embedded MTP head (21.1 GiB), q8_0 KV, 65k context	llama.cpp (Vulkan), llama-server	57.4	not stated	context: 65,536 context, q8_0 KV📏 llama-server request method (128 tokens, temp 0, seed 42, ignore EOS, no prompt cache), measured 2026-10-14, baseline without drafting; quiet power mode
Qwen3.6 35B-A3B	Unsloth UD-Q4_K_M with embedded MTP head (21.1 GiB), q8_0 KV, 65k context	llama.cpp (Vulkan), llama-server, MTP draft 3	81.9	not stated	⚡ spec-decode (MTP)context: 65,536 context, q8_0 KV📏 llama-server request method, measured 2026-10-14; +43% over 57.4 without drafting; code 88.6, prose 75.2; draft 2 = 80.1; quiet power mode
LFM2.5 24B-A2B	Unsloth UD-Q4_K_XL (13.9 GiB)	llama.cpp (Vulkan)	100.2	2,217 @ 512-token prompt est. TTFT @8k ≈ 3.7 s (calculated from measured prefill)	context: llama-bench tg128, low depth; 8 threads, -ub 2048📏 llama-bench pp512/tg128, fully offloaded, measured 2026-10-14; q8_0 KV 98.8; at 8K depth tg64 89.2 / pp2048 1652.4; quiet power mode

Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.

Bryan Lurer's llm-bench write-up (posted 1 Nov 2026, Gufo section dated 6 Nov), 'strix' host. Run in the quiet 54 W mode on purpose: the author measured balanced (85 W) at only +5.9% generation for +56% power. Gufo's prompt reading was about 28-30% under Gufo's own published table (author suspects kernel options not yet applied); generation matched it. Not listed: rows run on older llama.cpp builds from Feb-Sep 2026 (Qwen3.5, gpt-oss 120B, Gemma 4, Qwen3.8 27B 11.7 tok/s) and Nemotron Lightning (date not stated). Distinct from the GMKtec EVO-X3 Windows build (#31839).

Source: https://bryanlurer.com/posts/llm-bench/

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-x2-strix-halo-128gb-gufo-qwen38-flash-next-blurer.svg)](https://getrigbase.com/builds/gmktec-evo-x2-strix-halo-128gb-gufo-qwen38-flash-next-blurer/)

HTML

<a href="https://getrigbase.com/builds/gmktec-evo-x2-strix-halo-128gb-gufo-qwen38-flash-next-blurer/" rel="noopener"><img src="https://getrigbase.com/badge/gmktec-evo-x2-strix-halo-128gb-gufo-qwen38-flash-next-blurer.svg" alt="Measured on RigBase" height="20"/></a>

Preview: · 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 LFM2.5 24B-A2B? · Can Strix Halo (Radeon 8060S) run Qwen3.6 35B-A3B? · Can Strix Halo (Radeon 8060S) run Qwen3.8 Flash-Next?

Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).