{"slug": "qwen-image-2-1", "title": "Qwen-Image-2.1", "summary": "Alibaba's Qwen released Qwen-Image-2.1, a 7B-parameter image generation and editing model that combines a 32-layer single-stream DiT with a Qwen3-VL 8B text encoder and a 64-channel RGBA VAE, but ships under the non-commercial Qwen Research License rather than the Apache 2.0 license of Qwen-Image 1. The model generates and edits transparent RGBA images natively, accepts up to 10 reference images per edit, and outputs up to 2752x1536 at 40 inference steps, with day-one support in diffusers, ComfyUI, vLLM-Omni, SGLang, LightX2V, and stable-diffusion.cpp. Qwen's self-reported Qwen-Image-Bench score of 60.28 ranks 6th, behind GPT Image 2.5 at 67.01 and ahead of Nano Banana 2 at 59.83, with no independent evaluation yet; community-reported Q8 runtime memory is 15.6 GB VRAM total (7.3 DiT + 7.7 text encoder + 0.6 VAE).", "body_md": "# Qwen-Image-2.1\n\nenthusiast\n**Image generation and editing in one 7B model.** Qwen-Image-2.1 is a 32-layer single-stream DiT (7B parameters in the visual generation component) with a Qwen3-VL 8B text encoder and a 64-channel RGBA VAE at 16x spatial compression. It generates and edits transparent RGBA images natively (the Layered lineage folded in), takes up to 10 reference images per edit, supports local edits via circles, painted marks, or separate masks with identity preservation for people and products, and outputs up to 2752x1536 natively (default 2048x2048, 40 inference steps). Efficiency comes from mixed-granularity attention (token-level causal masking for text, chunk-level for image) plus prefix KV cache reuse across denoising steps.\n\n**License is the catch.** Qwen-Image 1 shipped under Apache 2.0; this release drops to the Qwen Research License: non-commercial use only, and any commercial purpose requires a separate license from Qwen ([\\[email protected\\]](https://tokenstead.ai/cdn-cgi/l/email-protection)). The site open-weights classification counts it as open, but research and evaluation are the only permitted uses out of the box. Budget for the commercial license before building a product on it.\n\n**Ecosystem, day one.** diffusers QwenImage21Pipeline, native ComfyUI with official workflows, vLLM-Omni (FP8, CUDA graphs), SGLang, LightX2V, and stable-diffusion.cpp; AMD ROCm and FlagOS multi-chip support are noted in the README.\n\n**What to treat as claims.** The Qwen-Image-Bench score (60.28, rank 6, behind GPT Image 2.5 at 67.01 and ahead of Nano Banana 2 at 59.83) is Qwen own benchmark - self-scored, no independent eval yet. The memory figure is community-reported, not official: a Q8 runtime at 15.6 GB VRAM total (7.3 DiT + 7.7 text encoder + 0.6 VAE) on RTX 3090-class cards, with diffusers enable_model_cpu_offload() path for smaller ones. Community reports of a periodic dot pattern in midtones exist but are unverified; treat content marking as unknown until confirmed.\n\n- 7.0B\n- qwen research\n- 🇨🇳 China\n- Sep 2026\n\nSave your hardware and every model page answers the real question: will it run on *your* machine, and how fast?\n\n[Join free - save your rig →](https://tokenstead.ai/login?return_to=%2Fonboarding)\n\n## Run it locally\n\nPer-quant memory needs and a static \"can you run it?\" reference - no rig entry required\n\n### The reference hardware\n\n22 reference configs, drawn in-house. Scroll for more.\n\n### Can you run it? - reference rigs\n\n| Rig | Q8_0 | \n|---|---|\n| NVIDIA Jetson Orin NX 16GB | tight | \n| Single GTX 1080 Ti (11GB) | offload | \n| 4x H100 80GB (320GB) | fast 472.4t/s | \n| NVIDIA DGX Station 748GB | fast 282.1t/s | \n| 8x RTX 3090 rack (192GB) | fast 264.1t/s | \n| 4x RTX 5090 (128GB) | fast 252.7t/s | \n| AMD Instinct MI300X (192GB) | fast 187.7t/s | \n| 4x RTX 4090 (96GB) | fast 142.2t/s | \n| 2x RTX 5090 (64GB) | fast 126.4t/s | \n| 2x RTX 3090 (48GB) | fast 66.0t/s | \n| Single RTX 5090 (32GB) | fast 63.2t/s | \n| RTX PRO 6000 Blackwell (96GB) | fast 63.2t/s | \n| Mac Studio M4 Ultra 192GB | fast 42.0t/s | \n| Mac Studio M4 Ultra 512GB | fast 42.0t/s | \n| Single RTX 4090 (24GB) | fast 35.5t/s | \n| MacBook Pro M5 Max 128GB | fast 23.6t/s | \n| Dual EPYC 9004 + 768GB DDR5-4800 | ok 16.3t/s | \n| DGX Spark 128GB unified | ok 9.6t/s | \n| Ryzen AI Max+ 395 128GB | ok 9.0t/s | \n| Jetson AGX Orin 64GB | slow 7.2t/s | \n| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | slow 7.2t/s | \n| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | slow 5.4t/s | \n\nFit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.\n\n## Download options\n\n## Or run it in the cloud\n\n        No per-token API provider pricing tracked for Qwen-Image-2.1 yet.\n        For flagship list prices, see the\n        [calculator](https://tokenstead.ai/calculator).\n      \n\n[See who runs Alibaba in production →](https://tokenstead.ai/adoption/alibaba)\n\n## Inference cost over time\n\nData accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.", "url": "https://wpnews.pro/news/qwen-image-2-1", "canonical_source": "https://tokenstead.ai/models/qwen-image-2-1", "published_at": "2026-09-21 04:05:16+00:00", "updated_at": "2026-09-21 04:26:40.229614+00:00", "lang": "en", "topics": ["generative-ai", "ai-products", "ai-tools", "large-language-models", "ai-infrastructure"], "entities": ["Qwen", "Alibaba", "Qwen-Image-2.1", "Qwen3-VL", "ComfyUI", "vLLM-Omni", "SGLang", "GPT Image 2.5"], "alternates": {"html": "https://wpnews.pro/news/qwen-image-2-1", "markdown": "https://wpnews.pro/news/qwen-image-2-1.md", "text": "https://wpnews.pro/news/qwen-image-2-1.txt", "jsonld": "https://wpnews.pro/news/qwen-image-2-1.jsonld"}}