Builds / AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU, gfx1151)

Strix Halo 128 GB (Sixunited AXB35): Qwen3.8 Flash-Next IQ4_XS at 68.8 tok/s on halogen 0.18.23, Qwen3.8 27B ROCmFP4-FAST at 38.6 with MTP on ROCmFPX

Last updated: 11 Nov 2026

Source: github.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source

GPU(s): AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU, gfx1151)
Amazon · eBay (used) · Newegg
VRAM / unified memory (GB): 128
CPU: AMD Ryzen AI Max+ 395 (32 threads)
Amazon · eBay (used) · Newegg
Motherboard: Sixunited AXB35 (BeyondMax Series), BIOS AMI v1.07
Amazon · eBay (used) · Newegg
RAM: 128 GB LPDDR5X-8000 (124 GiB visible to Linux, GTT pinned pages capped at 108 GiB)
Amazon · eBay (used) · Newegg
PSU: not stated
Other: Kernel 7.3.0-rc2, BIOS VGM carve 1 GB (GPU allocates from GTT); ROCmFPX llama.cpp fork main 120141f1f (build 12501) built from source, HIP, TheRock ROCm 7.15 nightly libraries; halogen-flash-server 0.18.23 (closed-source, gfx1151-only container from Peonist AI)
Reported total cost: not stated
Power: not stated

Measured speed (generation tokens/sec)

Model	Quant	Engine	tok/s	Prompt processing (tok/s)	Details
Qwen3.8 Flash-Next	IQ4_XS GGUF (UD bit map)	halogen-flash-server 0.18.23 (closed source), MTP draft head + prompt lookup, ctx 131072	68.8	1,603 (prompt length not stated)	⚡ spec-decode (MTP + prompt lookup)context: 131,072 context, 600-token outputs, reasoning off📏 median of 4 prompt types at temp 0 (prose 44.7, code 71.4, json 73.7, reasoning 66.1); 68.8 at temp 0.7 too; each figure median of 3 runs after a fresh server start, server's own timings; measured 4 Nov 2026
Qwen3.8 27B	ROCmFP4-FAST (julianmb/Qwen-3.8-27B-ROCmFP4-FAST-GGUF), q8_0 KV	ROCmFPX llama.cpp fork main 120141f1f (HIP), --spec-type draft-mtp, 4 draft tokens	38.6	353 @ 12,148-token prompt est. TTFT @8k ≈ 23.2 s (calculated from measured prefill)	⚡ spec-decode (MTP)context: 131,072 context, 600-token outputs, reasoning off📏 median of 4 prompt types at temp 0 (prose 24.5, code 41.6, json 47.1, reasoning 35.5), 36.5 at temp 0.7; each figure median of 3 runs, -np 1, server's own timings; measured 4 Nov 2026

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.

Owner's GitHub setup repo (README updated 4 Nov 2026). Same box, same harness: the 27B on upstream llama.cpp Vulkan UD-Q4_K_XL with MTP n=2 measured 25.7 tok/s (28 Sep), so the FP4 fork is about 50% faster on decode; the source says the FP4 quant, not the fork, is what makes it fast (the fork with UD-Q4_K_XL was slower). Flash-Next decodes 78% faster than the dense 27B and prefills 4.5x faster at ~12k because only ~6B parameters are active per token. halogen 0.18.11 measured 63.7 on 19 Oct and 66.3 on 4 Nov, so a few percent is day-to-day spread. Both engines are non-stock (one fork, one closed binary).

Source: https://github.com/sypherin/strix-halo-setup

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/strix-halo-axb35-qwen38-27b-rocmfp4-mtp-flash-next-halogen-sypherin.svg)](https://getrigbase.com/builds/strix-halo-axb35-qwen38-27b-rocmfp4-mtp-flash-next-halogen-sypherin/)

HTML

<a href="https://getrigbase.com/builds/strix-halo-axb35-qwen38-27b-rocmfp4-mtp-flash-next-halogen-sypherin/" rel="noopener"><img src="https://getrigbase.com/badge/strix-halo-axb35-qwen38-27b-rocmfp4-mtp-flash-next-halogen-sypherin.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 Qwen3.8 27B? · 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).