Can the Radeon 8060S (Strix Halo) 96 GB run Qwen-Image?
Download 16-bit (40.9 GB). With the model's working memory it needs about 43.7 GB, leaving 52.3 GB spare on 96 GB. Quality: the original weights.
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | qwen_image_bf16.safetensors 16-bit · Comfy-Org/Qwen-Image_ComfyUI | models/diffusion_models | 40.9 GB | Download → |
| Text encoder | qwen_2.5_vl_7b_fp8_scaled.safetensors Qwen2.5-VL 7B FP8 · Comfy-Org/Qwen-Image_ComfyUI | models/text_encoders | 9.4 GB | Download → |
| VAE | qwen_image_vae.safetensors Qwen-Image VAE · Comfy-Org/Qwen-Image_ComfyUI | models/vae | 0.3 GB | Download → |
| Total download · keep about the same free on disk | 50.5 GB | |||
Alternatives: Qwen2.5-VL 7B BF16 (16.6 GB). The 16-bit text encoder is a little more faithful; the smaller one is chosen here because it loads faster and needs less RAM. Sizes read from Hugging Face (2026-09-25). “Download” links start the file directly; the file name links are the same files the official ComfyUI workflows use.
Shared memory: about 56.5 GB of your 128 GB. On this machine the model, the text encoder, the VAE and the operating system all use the same memory. That leaves room, so nothing has to be swapped to disk. calculated
02Every Qwen-Image file on 96 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit ← SAFETENSORS · Comfy-Org | 40.9 GB | 43.7 GB | fits · 52.3 GB spare | the original weights | Hugging Face → |
| Q8_0 GGUF · city96 | 21.8 GB | 24.6 GB | fits · 71.4 GB spare | practically identical to the original | Hugging Face → |
| FP8 SAFETENSORS · Comfy-Org | 20.4 GB | 23.2 GB | fits · 72.8 GB spare | practically identical to the original | Hugging Face → |
| Q6_K GGUF · city96 | 16.8 GB | 19.6 GB | fits · 76.4 GB spare | very close to the original | Hugging Face → |
| Q5_K_M GGUF · city96 | 14.9 GB | 17.7 GB | fits · 78.3 GB spare | close; small differences in fine detail | Hugging Face → |
| Q4_K_M GGUF · city96 | 13.1 GB | 15.9 GB | fits · 80.1 GB spare | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_M GGUF · city96 | 9.7 GB | 12.5 GB | fits · 83.5 GB spare | noticeable loss of detail | Hugging Face → |
| Q2_K GGUF · city96 | 7.1 GB | 9.9 GB | fits · 86.1 GB spare | heavy loss; a last resort | Hugging Face → |
Sizes from the Hugging Face file listing, read 2026-09-25. “Needed” = file + 2 GB working memory for this model + 0.8 GB kept free for the system.
03The text encoder
Qwen2.5-VL 7B: 16.6 GB as 16-bit, 9.4 GB as FP8. ComfyUI encodes the prompt first and can push the encoder out of VRAM before sampling, so it does not have to fit together with the model. On 96 GB the FP8 encoder fits on its own, so prompt encoding stays fast.
04About this GPU
This is an integrated GPU that uses system memory. AMD states that up to 96 GB of the 128 GB can be set aside as graphics memory on Windows, which is why it can hold models no consumer graphics card can. The catch is speed: the GPU itself (40 compute units) is far smaller than a big desktop card, and its memory bandwidth is a fraction of one, so each step takes much longer. Treat the verdicts here as 'it fits', not 'it is fast'.
Radeon 8060S (Strix Halo) 96 GB: 96 GB LPDDR5X (shared system memory) · 256-bit · 256 GB/s · RDNA 3.5. Everything that runs on the Radeon 8060S (Strix Halo) 96 GB →
05What more VRAM would change
Nothing to gain for this model: the Radeon 8060S (Strix Halo) 96 GB already runs a top-quality file entirely in VRAM. More memory would only help with bigger images, longer clips or several models at once.
06Measured and reported results
| Label | Setup | Result | Peak VRAM | Date | Source |
|---|---|---|---|---|---|
| reported | Qwen-Image-2512 BF16 + 4-step Lightning LoRA · 1328x1328 · 4 steps “"workflow": "Qwen-Image-2512-BF16-4-Step-LoRA.json", ... "duration_seconds": 75.37661480903625” kyuz0 Strix Halo ComfyUI toolbox benchmark (Ryzen AI Max, ROCm); cold run incl. model load, flags --disable-mmap --gpu-only --disable-smart-memory --cache-none; resolution from ben | 75.38 s / image | — | 2026-02-13 | raw.githubusercontent.com → |
Reported results are other people's numbers, copied as published, with a link. Settings, drivers and ComfyUI versions differ, so compare them with care. Send yours.
07Questions
How much VRAM does Qwen-Image need?
Around 9.9 GB with the smallest file (Q2_K) and 23.2 GB with an 8-bit file (FP8), counting working memory and a small system reserve. The full 16-bit file needs about 43.7 GB.
Which Qwen-Image file should I download for the Radeon 8060S (Strix Halo) 96 GB?
16-bit (40.9 GB) from Comfy-Org/Qwen-Image_ComfyUI.
Is FP8 faster than GGUF on the Radeon 8060S (Strix Halo) 96 GB?
No. This GPU has no FP8 compute, so ComfyUI converts FP8 weights back before the maths. FP8 only saves memory here; Q8_0 GGUF is the closer-to-original 8-bit pick.
How much do I need to download for Qwen-Image on the Radeon 8060S (Strix Halo) 96 GB?
About 50.5 GB for the model file, text encoder and VAE listed on this page, and the same again free on disk. With a 96 GB card, 64 GB of system RAM or more is recommended.