ignis v0.9.1

Released 2026-10-28. The container image, the archives and the notes as written: the release on GitHub.

[!WARNING] These downloads are not GPU builds: use v0.9.2. This release’s Linux tarball, Windows zip and container images were built without the cuda feature, so the server reports ready and answers with the CPU mock’s placeholder text whatever the model (#408). v0.9.2 is the same code built with CUDA.

The Playground’s workspaces

A chat now works on files with the model: a workspace is a named set of files the user and the model share, born with its first file, kept by the page in the browser’s store and outliving the chat that made it (#170, #388-#399, #404, guide). ignis itself stores no file (ADR 0073).

• The model’s tools (guide): list_files, read_file, write_file, edit_file and delete_file.
  • Every change is a version, and a write over a change the model has not read is refused until it reads it.
  • read_file reads text as numbered lines, a PDF as its text, an audio file as its transcript, and a picture with vision when the loaded model sees pictures.
  • make_image lands its pictures in the workspace as made-N.png.
• Files in: what the prompt box attaches (text, PDF, audio, pictures) goes into the chat’s workspace; the panel uploads, records audio, and imports a zip.
• The panel: each file’s preview (an HTML site across files, Markdown, SVG and CSV rendered, audio), full screen for pages, text files and the diff; its history, the diff of any two versions, restore, hand edits and renames; the workspace exports as a zip.
• Bridges: pictures move between an image session and a workspace, and a transcript goes to a workspace as Markdown.
• One workspace at a time per chat: the picker binds a chat, and Move to… moves it to another workspace or leaves it with none until its next file; the model is told of the move once.

A system prompt that never changes

The memory titles and the workspace’s files no longer sit in the system prompt: they reach the model in a turn notice, a developer message after the history, kept with its turn and re-sent byte for byte (ADR 0072, guide). The system prompt and the tools block stay the same at every turn, so a chat’s prefix is reused instead of recomputed when a note or a file changes.

Upgrade notes

• The server is unchanged: this release changes the Playground only, which ships inside the binary.
• The Workspace switch is on by default (Settings, Tools). An older chat’s attachments move into a workspace at its first turn with the switch on, and are named in that prompt.
• An older chat’s first request after the upgrade recomputes its prompt once: its system prompt no longer lists the memory titles.
• Pictures in a chat’s history need a model with vision: after a switch to a model without it, such a chat is refused with vision_disabled, as a chat with attached pictures already was.

Decisions

• ADR 0072: the Playground’s system prompt never changes, and what changes rides after the history
• ADR 0073: a workspace is kept by the page, and ignis stores no files
• Recorded for work still open, not in this release: ADR 0074, a catalog entry may state its family, and an alias may be another entry’s id (#386).

Container image (linux/amd64)

podman run --rm --device nvidia.com/gpu=all -p 8000:8000 \
  -v /path/to/models:/models:ro \
  -e IGNIS_MODEL_ARTIFACT=/models/<name>.ninfer \
  ghcr.io/gpillon/ignis:0.9.1

(docker: --gpus all in place of --device.) The image carries the CUDA runtime but no driver — the host’s NVIDIA driver is injected by the container runtime. It is built from the same compile as the Linux tarball below, so the two always carry the same binaries.

Binaries

The kernel is compiled for SM120a (RTX 5090 / Blackwell consumer) only, and the engine needs a .ninfer artifact, which is not part of this release. The Linux tarball needs the CUDA 13 runtime on the host; the Windows zip carries cudart64_*.dll. Both need an NVIDIA driver new enough for CUDA 13.

Apache-2.0 (LICENSE, NOTICE); the vendored reference ops keep their own terms (NOTICE-kernel).

Full Changelog: https://github.com/gpillon/ignis/compare/v0.9.0...v0.9.1