Best local image generation models in November 2026: VRAM, speed and licenses

On 7 November 2026 the best open-weight image models are Qwen-Image-2.1 (Elo 1045), FLUX 3 Image Klein (1043) and Fibo 2.0 (1039), statistically tied at the top of Artificial Analysis's AA-Image-T2I v2.1 open-weights board and about 214 Elo behind the best closed model. None of the three is Apache 2.0: Qwen-Image-2.1 and Fibo 2.0 are non-commercial, and FLUX 3 Image Klein is free only for companies under $5M annual revenue. For a 32 GB card I would start with FLUX 3 Image Klein (11.1 GB of BF16 weights, about 4 seconds per 1 MP image on an RTX 5090 per BFL), or with Z-Image-Turbo if you need an Apache 2.0 license at any company size.

Elo from the Artificial Analysis AA-Image-T2I v2.1 open-weights leaderboard (FLUX.2 [dev] fixed at 1000; the page shows no update date), read 7 November 2026. Arena rank from the Arena text-to-image leaderboard, last updated 6 November 2026 (7,016,263 votes, 84 models). Licenses from each Hugging Face card or license file, checked 7 November 2026. Ranks 1 to 3 on AA overlap within their 95% intervals.

Which open image models rank highest in November 2026?

Three releases from September and October share the top, and all trail the closed leaders by about 200 points. On the overall AA-Image-T2I v2.1 board, OpenAI's GPT Image 2.6 Sunburst (max) leads at 1259, and the first open-weight model, Qwen-Image-2.1 at 1045, sits in ranks 21 to 27. The Arena board shows the same gap: GPT-Image-2.5 Sunburst scores 1429 and Qwen-Image-2.1 1227, a 202-point difference by my subtraction, and Arena still marks the Qwen score as preliminary.

Below the top the boards disagree. Arena ranks NVIDIA's Cosmos3-Super-Text2Image (29th in agentic mode) and HunyuanImage 3.0 (36th) above FLUX.2 [dev] (40th); AA puts FLUX.2 [dev] and Cosmos3 within 13 points. FLUX 3 Image Klein and Fibo 2.0 are too new for Arena. Stable Diffusion 3.5 Large, still a tutorial default, is 83rd of 84.

Which licenses allow commercial use?

Six of the 14 models in the table can be self-hosted commercially by a company of any size: the Apache 2.0, MIT and OpenMDW ones. The top three are all restricted.

• Permissive. Z-Image-Turbo, Qwen-Image-2512, ERNIE-Image and FLUX.2 klein 4B are Apache 2.0. HiDream-O1-Image and GLM-Image are MIT. Cosmos3-Super-Text2Image is OpenMDW-1.1, and its card says it is "ready for commercial and non-commercial use".
• Revenue-capped. FLUX 3 Image Klein's FLUX Kommunity License allows commercial use for individuals and companies under $5M annual revenue, per BFL's launch post; the license file adds that this is not blanket permission for commercial fine-tuning or redistribution. SD 3.5 Large is free below $1M revenue. Krea 2 uses its own Krea 2 Community License, whose thresholds I could not confirm on a first-party page.
• Non-commercial. FLUX.2 [dev] and FLUX.2 klein 9B use the FLUX Non-Commercial License v2.1, which bars "revenue-generating activity" with the model but lets you use outputs commercially. The Qwen Research License on Qwen-Image-2.1 is "for non-commercial purposes only". Fibo 2.0 is CC BY-NC 4.0 and gated. Ideogram 4.0's card says Ideogram 4 Non-Commercial, whatever Arena's "Ideogram Open Model" label suggests.

One trap for European readers: the Tencent Hunyuan Community License defines its territory as worldwide "excluding the territory of the European Union, United Kingdom and South Korea". Arena calls HunyuanImage 3.0 open, but the license grants nothing in the EU. This is not legal advice; read the license file before you ship.

How much VRAM does FLUX.2 dev need, and what fits in 32 GB?

FLUX.2 [dev] does not fit a 32 GB card at FP8; its 4-bit builds do. Hugging Face's FLUX.2 launch post says BF16 inference without offloading takes more than 80 GB and a 4-bit NF4 build runs in about 20 GB. The table below adds file sizes from the Hugging Face repos, which is the floor before activations.

File sizes in decimal GB from the Hugging Face file listings, checked 7 November 2026. VAE files (0.17 to 0.34 GB) are left out. The sum is my addition and ignores activations, which grow with resolution.

The rule I use: diffusion weights plus activations must fit; the text encoder only has to fit while it encodes the prompt, because ComfyUI can unload it before sampling. That is why FLUX.2 [dev] at NVFP4 (21.04 GB) works on an RTX 5090 even though weights plus encoder sum to 33.32 GB, while the FP8 build (35.46 GB for the transformer alone) has to stream weights from system RAM on every step. Measured peaks back this up. Stable Yogi logged 19.73 GB peak VRAM for Qwen-Image-2.1 at BF16 on an RTX 5090, about 14 GB of which is the 7B transformer at 2 bytes per parameter (my estimate). ComputingForGeeks saw peaks of about 11.2 to 15.3 GiB for FLUX 3 Image Klein on a 16 GB RTX 5060 Ti. BFL's card quotes 31 GB peak at BF16 with the unoptimized reference code on a B200, so the same model can look tight or roomy depending on the runtime.

Three of the bigger open models are out of reach on any consumer card. Cosmos3-Super-Text2Image has 64B parameters, so about 128 GB at BF16 by my calculation, and NVIDIA's card recommends an 8x H100 node and tests BF16 only. HunyuanImage 3.0 has 80B total parameters; Tencent's card asks for at least three 80 GB GPUs. FLUX.2 [dev] at full BF16 is about 64 GB of transformer by the same arithmetic. FP8 and 4-bit trade-offs are covered in NVFP4 vs MXFP4 vs FP8, and the GB10-specific steps are in the quantization docs.

How fast are they on an RTX 5090 or RTX 4090?

Models in the 5B to 7B range make a 1024x1024 image in 3.5 to 9 seconds on an RTX 5090 in the reports I found; FLUX.2 [dev] at FP8 takes over a minute because it spills out of VRAM. Settings differ by reporter, so compare rows with care.

Warm runs after model load unless noted. The Runflow author says a screen recorder on a second GPU changed run-to-run behavior, so treat that range as a sample. No source I found reports FLUX.2 [dev] NVFP4 seconds per image on an RTX 5090.

Two things in that table are easy to miss. NVFP4 is not automatically faster: on the RTX 5060 Ti, ComputingForGeeks measured FLUX 3 Image Klein at 7.8 seconds in NVFP4 against 6.0 seconds in INT8, and the author ties that to ComfyUI's current kernels. NVIDIA's March 2026 ComfyUI post claims 2.5x speed and 60% less VRAM with NVFP4 on FLUX.2 klein, tested on an RTX 5090, so the gain depends on model and kernel. And the RTX 4090 has no NVFP4 path: ComputingForGeeks ran NVFP4 only on its Blackwell card because the format needs Blackwell, so on Ada, FP8 or GGUF is the practical choice. Power is cheap to trim: Stable Yogi found 500 W only 3.6% slower than 575 W for Qwen-Image-2.1.

What does 1,000 images cost on a rented RTX 5090?

Under a dollar for the fast distilled models, about $2 for Qwen-Image-2.1, and $17 or more for FLUX.2 [dev] at FP8. My calculation is cost = seconds per image x 1,000 / 3,600 x hourly rate, at the $0.89 per hour that GPUwerk charges for its RTX 5090 (I run GPUwerk; any per-minute rental at a similar price gives the same result).

• Z-Image-Turbo, 3.5 s: 3,500 s = 0.97 h, so $0.87 per 1,000 images (about 1,030 per hour).
• FLUX 3 Image Klein INT8, 4 s: 4,000 s = 1.11 h, so $0.99 (900 per hour).
• Qwen-Image-2.1 BF16, 8.39 s: 8,390 s = 2.33 h, so $2.07 (429 per hour).
• FLUX.2 [dev] FP8 with offload, 68 to 145 s: 18.9 to 40.3 h, so $16.81 to $35.85.

Downloads, warm-up and rejected seeds are not included. BFL lists its hosted flux-3-image-klein API "from $0.009 per 1K image" without defining "1K image", so I would not build a break-even on it. The FLUX.2 [dev] line is the lesson: a model that spills out of VRAM costs 20 to 40 times more per image.

Can a DGX Spark run these image models?

Yes, and it loads models a 32 GB card cannot, but it is slower per image because diffusion is limited by compute, not bandwidth. In BFL's launch post, NVIDIA's numbers for the FLUX 3 Image Klein base model at 50 steps and 1024x1024 range from 1.7 seconds on a GB300 to 43.8 seconds on a DGX Spark. By step count, the distilled 16-step model should take about 14 seconds (my estimate), against the RTX 5090's reported 4. Forum timings for FLUX.2 klein, Z-Image-Turbo, SDXL and Qwen-Image-2512 on GB10 are collected in ComfyUI and Stable Diffusion on a DGX Spark.

The Spark's case is capacity. Its 128 GB of unified memory holds FLUX.2 [dev] at full BF16 (about 64 GB of transformer plus a 35.58 GB BF16 Mistral encoder, about 100 GB by my sum) with no offloading, or an image model next to an LLM. Cosmos3-Super at BF16 (about 128 GB) would fill nearly all of one Spark. The DGX Spark vs RTX 5090 comparison covers the general trade, and video diffusion, where memory pressure is worse, is in video generation on a DGX Spark.

Which image model should you run locally?

For one person on a 32 GB card with a business under $5M revenue, I'd pick FLUX 3 Image Klein: it ties for first among open models, the BF16 weights and encoder fit together in 20 GB, and it edits with up to 10 reference images. If the license has to be Apache 2.0 or MIT at any company size, the choice depends on what you value:

• Speed and small VRAM: Z-Image-Turbo (Arena 1082, 6B, about 8 GB at NVFP4).
• Quality with a permissive license: Qwen-Image-2512 (Arena 1125) at FP8 on a 32 GB card, or HiDream-O1-Image (Arena 1115, MIT).
• Research or personal work only: Qwen-Image-2.1, the top open model on both boards.
• Avoid in the EU: HunyuanImage 3.0, whose license does not apply there.

Skip FLUX.2 [dev] on a 24 or 32 GB card unless you use the NVFP4 build, and skip SD 3.5 Large for new projects. Arena Elo measures general preference, not your product shots or typography, so run your own prompts first. On GB10, the five checks for a new model apply to image models too.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• AA-Image-T2I v2.1 Open Weights Leaderboard, Artificial Analysis
• AA-Image-T2I v2.1 Leaderboard, Artificial Analysis
• Text-to-Image Arena leaderboard, Arena (formerly LMArena), 2026-11-06
• FLUX 3 Image Klein: open weights, 5.6B parameters, Black Forest Labs, 2026-10-15
• black-forest-labs/flux-3-image-klein, Hugging Face
• FLUX 3: One Multimodal Model, Black Forest Labs
• FLUX 3 Max: Text and Image to Video, Post-Trained by fal, fal, 2026-11-05
• FLUX.2 [klein]: Towards Interactive Visual Intelligence, Black Forest Labs, 2026-01-15
• black-forest-labs/FLUX.2-dev (model card and LICENSE.md), Hugging Face
• Diffusers welcomes FLUX.2, Hugging Face, 2025-11-25
• black-forest-labs/FLUX.2-dev-NVFP4, Hugging Face
• Comfy-Org/flux2-dev, Hugging Face
• Qwen/Qwen-Image-2.1 (model card and LICENSE), Hugging Face
• Qwen/Qwen-Image-2512, Hugging Face
• Comfy-Org/Qwen-Image_ComfyUI, Hugging Face
• Tongyi-MAI/Z-Image-Turbo, Hugging Face
• briaai/Fibo-2.0-dev, Hugging Face
• Ideogram 4.0 Technical Details, Ideogram, 2026-06-03
• ideogram-ai/ideogram-4-nf4, Hugging Face
• nvidia/Cosmos3-Super-Text2Image, Hugging Face
• tencent/HunyuanImage-3.0-Instruct (model card and LICENSE), Hugging Face
• HiDream-ai/HiDream-O1-Image, Hugging Face
• baidu/ERNIE-Image, Hugging Face
• stabilityai/stable-diffusion-3.5-large, Hugging Face
• krea/Krea-2-Raw, Hugging Face
• zai-org/GLM-Image, Hugging Face
• Qwen Image 2.1 RTX 5090 benchmark: 500 W was 3.6% slower, Stable Yogi, 2026-09-27
• FLUX.3 Klein in ComfyUI: 8GB VRAM Tested, ComputingForGeeks, 2026-10-24
• Best GPU for ComfyUI and Stable Diffusion, Benchmarked (2026), ComputingForGeeks, 2026-10-20
• GPU Benchmark Flux DEV fp8 5090 4090 3090 (Discussion #9002), Comfy-Org/ComfyUI on GitHub
• Z-Image: Powerful and highly efficient image generation model with 6B parameters, Hacker News
• Flux 2 in ComfyUI: install + workflow guide (2026), Runflow, 2026-06-18
• NVIDIA and ComfyUI Streamline Local AI Video Generation for Game Developers and Creators at GDC, NVIDIA, 2026-03-10