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

Can Strix Halo (Radeon 8060S) run Qwen3.8 Flash-Next?

Last updated:

Yes. Measured 36.1 tok/s generation (EngramHalo.cpp (llama.cpp fork, ROCm/HIP), Q4_K_XL + MTP (dzannotti head), KV q8_0; source).

Every measured result for Qwen3.8 Flash-Next on Strix Halo (Radeon 8060S)

BuildQuant / setupEnginetok/s
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 Flash-Next Q4_K_XL at 20.8 tok/s, 36.1 with MTP (EngramHalo.cpp)Q4_K_XL + MTP (dzannotti head), KV q8_0EngramHalo.cpp (llama.cpp fork, ROCm/HIP)36.1 CPU offloadsource
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 Flash-Next UD-IQ4_XS, 16.8 tok/s plain, 25-32 tok/s with MTP on ROCmUD-IQ4_XS + native MTP draft head + ngram-modllama.cpp ROCm (drluoto/strix-halo-flash-next branch), --spec-type draft-mtp,ngram-mod31.7source
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 Flash-Next UD-IQ4_XS, 16.8 tok/s plain, 25-32 tok/s with MTP on ROCmUD-IQ4_XS + native MTP draft head + ngram-modllama.cpp ROCm (drluoto/strix-halo-flash-next branch), --spec-type draft-mtp,ngram-mod25.4source
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 Flash-Next Q4_K_XL at 20.8 tok/s, 36.1 with MTP (EngramHalo.cpp)Q4_K_XL, KV q8_0, no MTPEngramHalo.cpp (llama.cpp fork, ROCm/HIP)20.8 CPU offloadsource
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 Flash-Next UD-IQ4_XS, 16.8 tok/s plain, 25-32 tok/s with MTP on ROCmUD-IQ4_XS, no speculationllama.cpp ROCm (drluoto/strix-halo-flash-next branch)16.8source

Memory: Strix Halo (Radeon 8060S) has 128 GB. Qwen3.8 Flash-Next weights are about 110.1 GB at Q4_K_M (180B params × 4.89 bits ÷ 8, before context/KV cache). Arithmetic, not a measurement; long contexts need more. Check other quants in the VRAM calculator.

Speeds are decode (generation) tokens/sec as reported by each source; quant, engine and context differ, so small gaps are noise.

All measured speeds for Qwen3.8 Flash-Next: best home rig for Qwen3.8 Flash-Next.

Qwen3.8 Flash-Next on other hardware: 2x DGX Spark (GB10) · 2x Intel Arc Pro B70 · 2x RTX 5070 Ti · 2x RTX 5080 · 2x RTX PRO 4000 Blackwell · 2x Tesla V100 32GB · 4x RTX 3090 · 6x RTX 3090 · Apple M5 Ultra · DGX Spark (GB10) · RTX 5090

Other models measured on Strix Halo (Radeon 8060S): Gemma 4 26B-A4B · Gemma 4 31B · Kolibri-1 · Qwen3.6 27B · Qwen3.6 35B-A3B · Qwen3.8 27B · gpt-oss 120B

Not enough VRAM? Rent before you buy. Run Qwen3.8 Flash-Next on a bigger GPU on a rented 24–96 GB cloud GPU by the hour and see the real speed before spending €1,000+. Try it on Runpod
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.