ignis v0.9.4

Released 2026-10-31. The container image, the archives and the notes as written: the release on GitHub.

The uncensored variant on 16 GB cards

The 27B’s uncensored variant now has a 16 GB build, the built-in catalog entry qwen3.8-27b-abliterated-16g (#386, guide). On a 16 GB card (RTX 5060 Ti 16 GB, 5070 Ti, 5080):

ignis-server model download qwen3.8-27b-abliterated
ignis-server --model-id qwen3.8-27b-abliterated

On a 16 GB card the id qwen3.8-27b-abliterated resolves to the 16 GB build, as qwen3.8-27b resolves to qwen3.8-27b-16g. A 32 GB or 96 GB card keeps the NVFP4 build under the same id, so a client names the same model on every card. It does not refuse: put the guardrails in the tool layer.

What it is: the 16 GB artifact with the 70 matrices of huihui-ai’s abliteration re-quantized at the bit width the 16 GB artifact gives each, by replaying ExLlamaV3 1.5.3’s conversion of those matrices. The other 1,152 of the 1,222 objects are the 16 GB artifact’s, byte for byte. Published at gpillon/Qwen3.8-27B-trellis-3bpw-abliterated-ignis, 13,379,877,124 bytes.
Its gates, registered before the run, measured on an RTX 5090 held to a 16 GiB budget (finding):
- KLD from the abliterated BF16: 0.0409 (95% CI 0.0351-0.0467), within 1.2× the 16 GB artifact’s 0.0403 from the base BF16;
- DFlash2 tokens per round: 0.999× the 16 GB artifact’s on the same recorded agent turns (95% interval 0.995-1.003, 83 of 116 jobs);
- Refusals: it answers the two probes the 16 GB artifact refuses.

The catalog

A catalog entry may state its model family, family: <the family's id>, and an alias may now be another entry’s served id of the same family, not only a family’s default id (#419, ADR 0074, operator catalog). The artifact’s header still has to agree, checked before anything stops. An operator catalog written without family loads unchanged.
The converter (tools/27b-16gb-converter) takes an override of EXL3 records on run and verify, made by its new replay of ExLlamaV3’s conversion (#420, #421). Two fixes from building with it: the reconstruction step empties the CUDA cache before each matrix, and keeps one shard open at a time, since on Windows every open shard is charged its whole size as commit.
ignis-artifact-inspect prints the artifact’s content hash, the digest that retained state is keyed by (#422).

Other changes

The Playground asks the browser for persistent storage at start, so a short disk cannot evict the sessions, and Settings › General says whether they are kept and offers to ask again (#427).
The telemetry tests that read a line logged after a request’s stream closed now wait for it, which removes the flake of the v0.9.3 tag’s Windows run (#424).

Upgrade notes

An operator catalog can no longer list qwen3.8-27b-abliterated as an alias: the built-in qwen3.8-27b-abliterated-16g answers to it now, and an id answers for one entry (ADR 0064).
make UNCENSORED=1 is unchanged: on a 16 GB card, start with ignis-server --model-id qwen3.8-27b-abliterated.

Decisions

ADR 0074: a catalog entry may state its family, and an alias may be another entry’s id
ADR 0047: a model catalog the operator can extend (amended)
ADR 0064: a catalog entry states the smallest card class it loads on (amended)

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:0.9.4

(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/v0.9.3...v0.9.4