cudabenchmarktest/Ornith-1.5-9B-Omni-GGUF

Ornith-1.5-9B-Omni (Q4KM)

This repository supplies the custom six-view GGUF sidecar for the logical Ollama model robit/ornith-1.5-omni:q4km. Its base reasoning model is the stock `ornith-ai/Ornith-1.5-9B`, not the OBLITERATUS derivative.

The release combines:

• —stock Ornith 1.5 9B text generation, native image vision, parsed optional thinking, and structured tools;
• —Qwen3-Omni audio, image, and sampled-video comprehension;
• —Qwen3-TTS text-conditioned 24 kHz speech generation.

These are independently executable graphs connected by a documented semantic adapter. This is not a claim that incompatible hidden states were tensor-spliced.

Published locations

• —Ollama: `robit/ornith-1.5-omni:q4km`
• —Ollama default alias: `robit/ornith-1.5-omni:latest`
• —Runtime: `robit-man/qwen-omni-adapters`
• —Build/release source: `robit-man/fine_tuning_suite`

Important: sidecar, not a standalone graph

ornith-1.5-9b-omni-q4km.gguf is one valid GGUF v3 container containing six namespaced model/projector views. It is not a standard single-architecture FROM target for stock Ollama.

The Ollama tag retains normal stock model/projector layers and adds the sidecar as application/vnd.robit.ollama.omni.bundle.v1+gguf. Stock Ollama executes text, native image vision, tools, and thinking. The Robit adapter resolves the same installed tag and runs the audio/video/TTS views with the pinned llama.cpp runtime.

Install and run

ollama pull robit/ornith-1.5-omni:q4km
git clone https://github.com/robit-man/qwen-omni-adapters.git
cd qwen-omni-adapters

OMNI_MODEL=robit/ornith-1.5-omni:q4km \
OMNI_LANGUAGE_MODEL=robit/ornith-1.5:9b \
./deploy.sh

The deployment validates the sidecar, materializes disposable media views, starts CUDA workers under the host GPU broker, exercises local smoke gates, and prints the authenticated portal URL. Ordinary Ollama clients may use the same tag directly for text/image/tools/thinking.

Advanced users can inspect or extract the Hugging Face sidecar with:

python -m training_suite omni-inspect ./ornith-1.5-9b-omni-q4km.gguf
python -m training_suite omni-attach your-exact-stock-ornith-tag:q4km \
  ./ornith-1.5-9b-omni-q4km.gguf
python -m training_suite omni-prepare your-exact-stock-ornith-tag:q4km \
  --out ./runtime-cache

The target must contain the exact base and projector digests below. Attaching the sidecar to an arbitrary model does not make that model Omni.

Artifact inventory

The base namespace is unprefixed; b.p.* is its projector; a.c.m.* and a.c.p.* are comprehension; s.t.m.* and s.t.p.* are TTS. The exact machine-readable inventory is published as ornith15-stock-sidecar-manifest.json.

Capability boundary

The model-level generation setting uses Ollama's num_predict=-1, removing the source tag's inherited 16,384-token output cap. Clients may still set a per-request output limit.

Limitations and safety

• —Stock Ollama does not execute audio/video/TTS from the custom layer; use the adapter for those routes.
• —The bridge is semantic text, not a learned dense connector.
• —Video support is comprehension only, not video generation.
• —Media, transcripts, OCR, captions, and tool results are untrusted input.
• —Public deployments need authentication, rate limits, isolated tool execution, bounded decoding, monitoring, and appropriate output controls.

See the included release record, sidecar manifest, validation report, and third-party notices for reproducibility and attribution.