Pictures from the same card

The card that runs your coding agent can also draw. ignis serves Qwen-Image-2.1-Turbo on POST /v1/images/generations and POST /v1/images/edits, the routes OpenAI’s SDKs already speak, so a client that makes pictures needs no second server. A 2048x2048 image takes 7.87 s end to end on the 5090, streamed through the Playground’s own calls.

Non-commercial licence

The image weights are under the Qwen Research License: non-commercial use only, with a “Built with Qwen” notice. Do not serve the image model in a commercial product. The licence, and where the weights live, are on the Images page.

How it works

The image model makes pixels, so lanes, KV and generation caps do not apply to it. It is reached by a switch from a text model: the card holds one model at a time, so the 27B unloads and the image model loads (see Models that come and go). The kernels are ignis’s own for the 5090: an FP8 W8A8 GEMM and an 8-bit attention for the 8-step denoising loop, a VAE decode, and a condition encoder for edit inputs. An edit takes one to ten PNG, JPEG or WebP images, and stream: true sends progress events by phase (enhancing, switching, loading, encoding, denoising, decoding).

After the weights, the VRAM that is left sets a condition budget of prompt and input-image tokens; larger inputs are scaled down, not refused. prompt_enhancement has the 27B rewrite the prompt first with Qwen’s own system prompts; with the 27B already loaded, the rewrite runs first and the server switches once. The Playground’s Image studio adds image sessions, marks drawn on a picture, and a make_image tool the chat can call: it switches to the image model, makes the pictures and switches back. The pictures it makes land in the chat’s workspace as files, and the model reads them back with read_file when the loaded model has vision (tools and agents).

The numbers

Figure	Measured on	Source
The denoising loop: 1.20 s at 1024x1024 (145-152 ms a step) and 6.11 s at 2048x2048 (747-802 ms a step); whole jobs 1.37 s and 6.79 s	in-process, 8 steps, RTX 5090	Image model on the 5090
Edit at 2048x2048: 7.62 s from one image, 10.72 s from four	in-process	same
The FP8 GEMM runs at 67-73% of the literature’s FP8 bar (498-507 TFLOPS against 690-740); the 8-bit attention at 88-93% of its bar	RTX 5090	same, and Image model kernels
Served end to end, streamed: a 2048x2048 generation 7.87 s (denoising 5.91 s of it); a 2-image edit 9.48 s	the Playground’s own calls, RTX 5090	Image serving spec, as built, the live run of 2026-10-14
Switching 27B to the image model 10.0 s, and back 11.3 s	same run	same

The ceiling is the GEMM, not the design

The kernels target SM120a, and the 7.87 s is a 5090 figure. The GEMMs run at about 70% of the literature’s bar, so they set the ceiling. Same-seed edits differ run to run in grain; read Image edit, same seed, grain before promising reproducible edits.

See it, control it

• Routes: POST /v1/images/generations and POST /v1/images/edits, model id ignis-image-2.1-turbo, fetched with ignis-server model download ignis-image-2.1-turbo.
• Sizes: image.sizes, every side a multiple of 32.
• Queue and inputs: image.queue_depth, image.max_n and image.max_condition_images.
• Prompt rewrite: image.prompt_enhancement.
• Metrics: ignis_image_jobs_total (by outcome), ignis_image_images_total, ignis_image_job_duration_seconds, ignis_image_job_phase_seconds_total and ignis_image_queue_jobs, in the metrics.
• Playground: the Image studio; image sessions are stored in the browser.

Limits

• One image job runs at a time; n images are made one after another.
• The response is an RGBA 8-bit PNG as b64_json. response_format: url and other output_format values are a 400.
• The image model is never co-resident with a text model: a chat while it is loaded pays a switch of 10-11 s, and the switch drops the old load’s retained state.

Read more

• Images: routes, sizes, the licence, streaming and the condition budget.
• Models that come and go: the switch an image job rides on.
• Image studio: the Playground’s image sessions.