I want to run it
Install the image or the archive, fetch a model, open the Playground and chat. No configuration needed.
The user path →
ignis
A CUDA inference engine for Qwen3.8 on Blackwell cards. OpenAI-compatible API, a Playground in the browser, one binary.
What sets it apart
Section titled “What sets it apart”One model family on one class of card, built for the load an agent makes: a main conversation and its subagents on the same GPU at once.
367 tok/sone lane, coding prompts, greedy1,065 tok/seight lanes at once, aggregate9,717 tok/scold prefill of a 46K prompt~100 tok/sa 125B MoE on a 32 GB card, one lane7.11xthe KV capacity of BF1636.5 msa decision, zero tokens decoded896Ktokens of context, with vision and DFlash27.9 sa 2048×2048 image, end to end
Why ignisSeventeen topics, each with how it works, what was measured and where it stops: the decode round, DFlash2, KV at 9 KB a token, retained state, decisions without decoding, reasoning that finishes.
The Playground, view by viewChat with subagents on their own lanes, Decide, the image studio, the live Monitor and the Server view, with screenshots.
The models, on Hugging Face
Section titled “The models, on Hugging Face”Every model the server fetches by itself is a public Hugging Face repository, pinned file by file at a fixed revision (Get a model).
Qwen3.8-27B, NVFP4 + DFlash2qwen3.8-27b, the default: 18.1 GiB, with its grafted DFlash2 drafter. RTX 5090 and RTX PRO 6000.
Qwen3.8-27B for 16 GB cardsqwen3.8-27b-16g: the 27B with a 3-bit trellis body, 12.5 GiB, for the RTX 5060 Ti 16 GB, 5070 Ti and 5080.
Qwen3.8-27B, uncensoredqwen3.8-27b-abliterated: the same 27B, abliterated (uncensored), 18.1 GiB.
Qwen3.8-Flash-Next, 125B MoEqwen3.8-flash-next: experts in a 2.5-bit trellis code, 67.8 GiB with its MTP head, on a 32 GB card.
Qwen-Image-2.1-Turboignis-image-2.1-turbo: the image model, 15.6 GiB. Non-commercial licence.
Pick your path
Section titled “Pick your path”Every page of this guide belongs to one site; each path is an order to read it in, from the first command to the deepest knob.
I want to tune it
Context length, VRAM and the KV pool, speculation, host RAM, KV-disk, profiles, air-gapped machines, Knative.
The advanced path →I want to build on it
The OpenAI-compatible API, the Responses API, streaming, tools, deadlines, model switching, images and decisions.
The developer path →I want to work on it
Build from source, the tests and the GPU discipline, make, specs, ADRs and findings, and how a change ships.
The contributor path →I am an AI agent
Working for a user: what ignis is, the checks before installing it, the choices that are the user’s, starting it and connecting a client.
The agent path →Jump in
Section titled “Jump in”InstallThe container image and the release archives.
Install with a coding agentTell the agent you already use to install it: what it checks and what it asks you.
Get a modelThe catalog, model download, and pointing at your own artifact.
ConfigurationEvery setting is one field: flag, env var, config file.
Every config fieldThe generated reference, straight from the server's field table.
TroubleshootingA start refused, a slow decode, a busy GPU: what each message means.
ChangelogEvery release, newest first, with its notes as published.