Qwen3.8-27B-E03-Obliterated-Omni

Adapters and Deployment package

This release supplies the custom GGUF sidecar used by the single logical Ollama model robit/qwen3.8-27b-e03-obliterated-omni:q4km. It combines:

• Qwen3.8-27B-E03-Obliterated language generation, parsed thinking, structured tools, and the original Qwen3.8 image-vision projector;
• Qwen3-Omni audio, image, and sampled-video comprehension;
• Qwen3-TTS text-conditioned 24 kHz speech synthesis.

The capabilities were tested through the exact sidecar installed in the local Ollama tag. This is semantic orchestration between independently executable graphs, not a claim that incompatible hidden states were tensor-spliced.

Important: custom sidecar, not a standalone model

qwen3.8-27b-e03-obliterated-omni-q4km.gguf is a valid, contiguous GGUF that contains six namespaced model/projector views. It is not a single standard GGUF architecture and must not be used as FROM in a stock Ollama Modelfile.

The published Ollama tag contains normal model/projector layers plus this custom media type:

application/vnd.robit.ollama.omni.bundle.v1+gguf

Unmodified Ollama executes text, images, thinking, and tools. The Robit Omni Adapter resolves the custom layer from that same installed tag, materializes disposable runtime views, and routes audio/video comprehension and TTS.

• Adapter overview
• Runnable examples
• Wire protocol
• Runtime guide
• Exact release record

Recommended installation

One Ollama pull installs the stock execution layers and the sidecar:

ollama pull robit/qwen3.8-27b-e03-obliterated-omni:q4km
git clone https://github.com/robit-man/fine_tuning_suite.git
cd fine_tuning_suite

python -m training_suite omni-resolve \
  robit/qwen3.8-27b-e03-obliterated-omni:q4km
python -m training_suite omni-prepare \
  robit/qwen3.8-27b-e03-obliterated-omni:q4km \
  --out ./runtime-cache

omni-prepare reconstructs the comprehension and TTS GGUF pairs and verifies them against the hashes embedded in the sidecar. Start the pinned llama.cpp workers and unified adapter using the runtime guide above. The materialized files are a disposable cache and should be removed after the workers stop.

Sidecar-only installation from this repository

Advanced users who already have the compatible stock Qwen3.8 base tag may download the GGUF here and attach it:

python -m training_suite omni-inspect \
  ./qwen3.8-27b-e03-obliterated-omni-q4km.gguf
python -m training_suite omni-attach \
  your-local-compatible-base:q4km \
  ./qwen3.8-27b-e03-obliterated-omni-q4km.gguf

The base tag must use the exact model and projector recorded below. Attaching the sidecar to an arbitrary architecture does not make that architecture Omni.

API examples

The adapter extends non-streaming POST /api/chat while retaining Ollama’s normal model, messages, tools, think, format, and options fields. Binary inputs and output WAV data use strict padded base64 JSON envelopes.

# 16 kHz mono PCM16 WAV → exact transcription
python examples/omni_adapter/client.py \
  --model robit/qwen3.8-27b-e03-obliterated-omni:q4km \
  asr ./speech-16khz-mono.wav

# MP4/WebM frames, optionally with its demuxed audio track → description
python examples/omni_adapter/client.py \
  --model robit/qwen3.8-27b-e03-obliterated-omni:q4km \
  video ./events.mp4 --fps 2 --max-frames 96 --include-audio

# Text → tagged base64 24 kHz mono PCM16 WAV
python examples/omni_adapter/client.py \
  --model robit/qwen3.8-27b-e03-obliterated-omni:q4km \
  --output-audio ./speech.wav \
  tts "Read this sentence."

The complete request/response shapes, media limits, tool-call speech deferral, and normalization rules are normative in the wire protocol. Adapter v1 is turn-based and requires "stream": false.

Capability and validation matrix

The media workers were started under scoped CUDA reservations. All release workers were stopped and all reservations released after validation.

Artifact inventory

The GGUF tensor namespaces are unprefixed for the base model, b.p.* for the base projector, a.c.m.* / a.c.p.* for comprehension, and s.t.m.* / s.t.p.* for TTS. See sidecar-manifest.json for machine-readable source revisions, byte sizes, tensor counts, and SHA-256 values for all six views.

Component provenance

The base derivative and both official upstream Qwen component families are published under Apache-2.0. Qwen is the work of the Qwen team; the E03 refusal-direction intervention is the work of its source author; the GGUF conversions are credited to their respective ggml-org repositories. This combined package is an independent release and is not an official Qwen, llama.cpp, Ollama, or source-author release.

See THIRD_PARTY_NOTICES.md for source links and attribution. The source model pages and their license files remain authoritative.

Operational requirements

• The Hugging Face sidecar is 38.8 GB. The full Ollama tag also carries normal model/projector/template layers.
• Preparing the four media-runtime views temporarily requires about 21.4 GB of additional disk space.
• GPU memory depends on context and runtime settings. The Q4_K_M language model is about 16.5 GB; the comprehension pair is about 19.9 GB; the TTS pair is about 1.5 GB before runtime overhead.
• Use explicit GPU placement/reservations when workers share a host. Never let an independently supervised CUDA worker allocate an anonymous device.
• Treat decoded media and semantic observations as untrusted input. Enforce byte, frame, duration, codec, context, timeout, and concurrency limits.

Limitations

• Stock Ollama does not internally execute audio input, video input, or audio output from this sidecar. Those routes require the linked adapter and llama.cpp workers.
• This does not add video generation. Video support is comprehension of sampled frames plus, when requested, a separately demuxed audio track.
• The cascade passes an explicitly delimited semantic observation into the Qwen3.8 language model; it does not preserve dense Omni hidden states.
• The reference TTS worker is serial and starts llama-tts per request. It is a validation implementation, not a high-concurrency speech service.
• Video/audio association is temporal but not sample-accurate.
• Adapter v1 does not stream input or output audio.

Safety

The E03 language checkpoint deliberately reduces refusal behavior and must not be treated as a safety-aligned replacement for upstream Qwen3.8. Its source reports known harmful-prompt repetition/breakage and limited qualification. Public-facing use should add access control, sandboxed tool execution, media decoding isolation, output filtering, rate limits, monitoring, and human review. Never expose privileged tools solely because the model emitted a valid call.

Reproducibility and cleanup

The build/release runbook defines the source-pinning, GGUF round-trip, live capability, registry round-trip, and cleanup gates. Do not delete source weights or materialized views until the repository, Hugging Face file, and Ollama tags have all been remotely verified. Afterward, stop workers and remove only the exact disposable view cache; it can be reconstructed from this sidecar.