Introducing Qwen-Image-2.1: image generation and editing in one model, with native transparency and a compact 7B visual generation component.🚀
🤖 https://t.co/ZaQGRRtyyk
🎨 Try it in Studio: https://t.co/NRAbnNBRb9
⚡ Efficient inference: KV cache reuse speeds up generation and editing while reducing memory usage, especially with multiple reference images.
🖼️ Native transparency: Generate regular or transparent images, edit transparent layers, and extract subjects from photos.
🛠️ Versatile editing: Combine up to 10 reference images, target local edits, and preserve portrait identity and product details.
✨ Refined aesthetics: Improved typography, portrait lighting, and realistic textures bring finer detail to generated images.
🤯 A localmaxxer hit ~381 tok/s on a SINGLE RTX 3090 with Qwen3.8-27B.
This developer has turned a 24GB RTX 3090 into a monster Qwen inference box - w/some creativity.
Four days ago 👉 ⚡ ~82 tok/s single-user
Then 👉
⚡ ~114 tok/s with optimized MTP
⚡ ~138 tok/s with DFlash2 + lookup drafting
Now 👉 🔥 ~381 tok/s on ONE request
How? The recipe combines ...
🧠 Qwen3.8-27B
🎮 1× RTX 3090 24GB @ 250W
⚙️ heavily optimized vLLM
⚡ DFlash2 speculative decoding
🔎 lookup-augmented drafting
📚 prefix caching
🧮 16-token verification blocks
💾 quantized KV / heads / activations
DFlash2 normally proposes 7 tokens.
The developer realized the verification block doesn't have to stop there.
If Qwen is answering from a document already sitting in the prompt, the system can fill the remaining draft positions using tokens found directly in that context.
🎯 So the target model can verify 16 tokens at once.
On a ~25K-token document reproduction task:
Previous DFlash2 👉 ~260 tok/s
Longer verification + context lookup 👉 🔥 ~382 tok/s
Acceptance:
🤯 15 of 16 tokens per verification step
That is where the crazy number comes from.
⚠️ On ordinary real-world chat prompts, the same setup is around ~133 tok/s
Still extremely fast for a dense 27B model on an RTX 3090.
The 381 tok/s mode shines when the answer largely comes from material already in context so these are best use cases
📚 RAG / document Q&A
💻 Coding assistants applying edits
📝 Quoting or rewriting documents
🔎 Extracting information from long prompts
And another optimization 👉 With prefix caching, a second question against the same 25K-token document reportedly goes from:
🐌 22.4 sec TTFT
→
⚡ 0.56 sec TTFT
Because the model doesn't need to process the whole document again.
🎯 It's specifically a mode for RAG front ends and coding agents.
Follow iamMess on Reddit or syv-ai on GitHub
🔗 Reddit: r/LocalLLaMA/comments/1vtup5s/
🔗 GitHub: /syv-ai/qwen38-27b-rtx3090
A single RTX 3090 hit 381 tok/s on Qwen3.8-27B.
DFlash2 + lookup: 138
Optimized MTP: 114
context lookup so the model verifies 16 tokens at once when the answer already lives in the prompt.
Longer verification & prefix cache: 381 tok/s
The 381 number only appears when the model is reproducing or extracting from a long prompt.
That’s the point.
Prefix cache also drops TTFT from 22s to 0.56s on the second question against the same document.
Best for RAG, document Q&A, and coding agents. Regular chat still lands around 133 tok/s.
25K-token context.