Skip to content Skip to the checker
RigCheckWhat can I run?

Qwen-Image-2.1 hardware requirements: VRAM, GPU and RAM

  • Page updated
  • Sources checked
  • Primary sources checked October 12, 2026
  • File sizes, published recipes and RigCheck assumptions labeled separately
  • Jump to the checker ↓

Qwen-Image-2.1 generates and edits images. Turbo uses the same 7B visual generation architecture with an accelerated eight-step checkpoint. The workflow also needs a Qwen3-VL-8B text/vision encoder and a model-specific VAE.

Short answer: ComfyUI BF16 files total 32.44 GB; INT8 ConvRot files total 17.28 GB, including the encoder and VAE. Base GGUF Q4 plus encoder and VAE is 10.02 GB, or 11.19 GB with the separate vision projector for editing. These are file totals, not VRAM minima: attention, reference inputs, prefix caches and VAE decoding add memory. Turbo's eight steps do not shrink its weights. Base CUDA ComfyUI and official Diffusers routes are documented; Turbo ComfyUI sampling remains unverified here. Base GGUF has a stable-diffusion.cpp route. Mac memory budgets remain conditional; no verified device measurement here. The Qwen Research License does not grant commercial use; commercial generation services need separate permission.

Plan a complete Qwen-Image workflow

Open in a full page ↗

Loading private checker…

Spec sheetSources checked
Developer
Qwen Official
Params
7B visual generation architecture Official
Components
Qwen3-VL-8B text/vision encoder and a model-specific VAE Official
Files
ComfyUI BF16 files total 32.44 GB Community
License
Qwen Research License Official
ComfyUI
Base CUDA ComfyUI and official Diffusers routes are documented; Turbo ComfyUI sampling remains unverified here Community
stable-diffusion.cpp
Base GGUF has a stable-diffusion.cpp route Community

Files and formats: the full workflow

Decimal GB from pinned Hugging Face file lists, checked October 12, 2026. The interactive file list names exactly what is counted; ComfyUI text encoders include vision weights. Optional prompt rewriters, ControlNet and LoRAs are excluded and need additional memory.

RouteDenoiserText/vision encoderVAE / projectorTotal files
ComfyUI BF16, base or Turbo14.230 GB17.534 GB0.676 GB32.440 GB Comfy-Org files
ComfyUI INT8 ConvRot, base or Turbo7.257 GB9.351 GB0.676 GB17.283 GB Comfy-Org files
Base GGUF Q4 generation4.200 GB5.149 GB UD-Q4_K_XL0.676 GB10.024 GB Unsloth + Comfy-Org
Base GGUF Q4 editingSameSameAlso 1.163 GB BF16 mmproj11.186 GB Unsloth + Comfy-Org
Official Diffusers BF16 base / TurboDifferent repository layout and stored VAE dtype; use the complete official checkpoint33.116 / 32.440 GB Qwen files

INT8 and GGUF are real community distributions, not theoretical parameter multipliers. Unsloth also publishes FP8 files. The checker uses Comfy-Org INT8 pairs and base GGUF only; it does not treat these quantizations as equivalent in image quality. Community Turbo GGUF files exist, but a correct Turbo runtime/sampling route has not been verified here. A base GGUF file is not Turbo. These image GGUFs use stable-diffusion.cpp rather than a chat-model Ollama or llama.cpp command.

ComfyUI, Diffusers and Mac routes

  1. ComfyUI, CUDA: update ComfyUI and follow Comfy-Org's native templates and folder placement. Put denoiser, encoder and VAE into their respective model folders. Keep refine_prompt off: the templates contain an optional Qwen3.5-9B prompt-enhancer branch excluded from these budgets. If validation requests its file, remove/bypass that branch or download the additional enhancer and budget it separately (about 9.47 GB INT8). These linked templates use base weights and a base sampler. Comfy-Org publishes full Turbo weights, but the template does not load Turbo’s saved Diffusers sigma sequence. Replacing the model and setting eight steps alone is insufficient; this page has not verified a ComfyUI Turbo sampler workflow. Use the complete official Diffusers checkpoint for Turbo’s documented route. Its acceleration LoRA is an alternative to the full checkpoint, with its own sampling requirements.
  2. Diffusers, official BF16: follow the Turbo quick start or base card. Turbo requires Diffusers support for checkpoint-configured sigmas (PR #14950); the card installs from GitHub and transformers ≥5.17.0. Use QwenImage21Pipeline. Turbo loads its saved eight-step schedule and CFG=1: setting num_inference_steps alone does not override the schedule. The checker does not budget ComfyUI files as an official Diffusers checkpoint.
  3. Base GGUF, CPU / CUDA / conditional Mac: follow the stable-diffusion.cpp Qwen Image 2.1 guide and its backend build instructions. Keep the 2.1 VAE; older Qwen Image or Wan VAE files are not interchangeable. Editing with a GGUF text encoder needs --llm_vision and the separate projector. Repeat -r for multiple references. Unsloth lists Apple Silicon and CPU GGUF memory estimates in its guide; they are estimates, not a measured Mac configuration. No exact Mac speed or successful allocation is promised.

Resolution, editing, offload and out-of-memory failures

Larger images and more references increase visual tokens, attention work, saved prefix context and VAE intermediates. Qwen supports up to ten reference images; actual reference dimensions and runtime preprocessing matter. Our slider changes only an explicit planning allowance, not a calibrated scaling law. There is no chat-token context slider or chat KV estimator.

Unsloth estimates 12–16 GB VRAM for GGUF Q4 at 1024², batch one, and lists INT8/FP8 at 512² on a 24 GB card; these are source-author recommendations, not RigCheck measurements. Do not transfer the small-canvas recommendation to 2048² or a multi-reference edit. Turbo reduces denoising passes, not the encoder, VAE or checkpoint memory, and its total speedup depends on encoding, decoding and transfer overhead.

If loading runs out of host RAM, verify all components and use the runtime's documented offload/loading options. If denoising fails, reduce resolution and references, use supported memory-efficient attention or offload; if decoding fails, use documented VAE tiling/offload. Offload trades RAM and transfer time for VRAM and can be slow on CPU or a constrained interconnect. No seconds-per-image estimate is shown without a matching measured setup. stable-diffusion.cpp's image-prefix cache is a separate image-model optimization; its guide explains cache dtype and retry behavior.

Research license and commercial services

The official Qwen Research License permits noncommercial research/evaluation; commercial use requires a separate license from Qwen (contact in the official license). This includes planning a paid hosted generation service. Community quantization does not remove the underlying terms. Read the official license before using the weights.

Measured Qwen-Image-2.1 on your own hardware, or tried and it failed? Send us the numbers: paste the benchmark output or log into the form (or email it). After checking it we add it to this page, labeled as a community report and shared under CC BY 4.0.

Common setups

The checker's own answer for 8 common setups (graphics cards with 32 or 64 GB of system RAM, then Macs), in the finder's quick mode (graphics cards: base model, INT8, 512×512, batch one, no reference images; Macs: base model, the GGUF files (not INT8), 512×512, batch one, no reference images). Answers are planning budgets, not measured setups; "Unknown" means the checker has no route or source for that hardware. Your own setup: the checker.

Qwen-Image-2.1: answers by setup
SetupAnswer
GeForce RTX 3060 (12 GB) + 32 GB RAM! Above the all-components-resident planning budget; offload peak unverified (planning budget)
Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction.
GeForce RTX 4060 Ti (16 GB) + 32 GB RAM– Unknown
No published route checked here for this device.
GeForce RTX 3090 / 3090 Ti (24 GB) + 64 GB RAM! Above the all-components-resident planning budget; offload peak unverified (planning budget)
Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction.
GeForce RTX 5090 (32 GB) + 64 GB RAM! Within the all-components-resident planning budget; allocation unverified (planning budget)
Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction.
Apple Silicon Mac (64 GB unified)! Mac GGUF is within the planning budget; Metal execution unverified here (planning budget)
Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction.
Mac with M5 Max, 40-core GPU (128 GB unified)! Mac GGUF is within the planning budget; Metal execution unverified here (planning budget)
Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction.
Mac Studio M3 Ultra (256 GB unified)! Mac GGUF is within the planning budget; Metal execution unverified here (planning budget)
Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction.
Mac Studio M3 Ultra (512 GB unified)! Mac GGUF is within the planning budget; Metal execution unverified here (planning budget)
Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction.

Machine-readable: the same answers as JSON (CC BY 4.0, attribution "RigCheck").

Related hardware

Graphics card and Mac pages where this page's checker gives an answer, smallest memory first: graphics cards with 64 GB of system RAM, Macs with their unified memory, in the finder's quick mode (planning budget; each page lists every model).

All graphics cards · All Macs · Any other machine

FAQ

How much VRAM does Qwen-Image-2.1 Turbo need?

No single verified minimum covers every workflow. Turbo ComfyUI sampling is unverified here; the official Diffusers route uses the saved eight-step sigma schedule. ComfyUI INT8 files total 17.28 GB, BF16 32.44 GB including encoder and VAE; offload can reduce residency, while resolution, references and decoding add memory. Turbo's eight steps do not reduce weight size.

Can I run Qwen-Image-2.1 on a Mac?

Base GGUF has a cross-platform stable-diffusion.cpp route and Unsloth lists Apple Silicon estimates. We have no verified run for your exact Mac. The checker reports a conditional planning budget, without a speed or an allocation guarantee; ComfyUI INT8/BF16 Mac compatibility is not inferred from memory.

Is there a Qwen-Image-2.1 GGUF?

Yes. Base Q4_K_M needs a text encoder and the 2.1 VAE too. Editing with a GGUF encoder also requires its vision projector. Community Turbo GGUF artifacts exist, but this page does not yet verify a Turbo GGUF sampling route.

Can I sell a Qwen-Image generation service?

The research license does not grant commercial use. Obtain a separate Qwen commercial license first; read the current official terms.

Sources