cudabenchmarktest/Qwen3.8-27B-E03-Obliterated-Omni-Audio-Bridge-GGUF

Qwen3.8 27B E03 Obliterated Omni Audio Bridge

This is an audio-bridge release for the `qwen-omni-adapters` runtime. It keeps manitcor/Qwen3.8-27B-Obliterated-E03 as the sole language trunk, retains the target's native vision path, reuses the frozen Qwen3-Omni audio tower, and replaces only the final 1,280-wide audio projection with the trained bridge. Qwen3-TTS remains an independently executed output graph in the lightweight sidecar.

The release combines multiple upstream components. Review each component's license and acceptable-use terms before redistribution or deployment.

Reduced operational weight set

The three deployable weight artifacts total 19,682,103,136 bytes (18.33 GiB). The previous full-router Ollama bundle occupied 56,321,596,306 bytes (52.45 GiB); the release weight set is 65.1% smaller.

This is an exact artifact/resident-weight comparison, not a process-peak claim. KV cache, graph workspaces, CUDA allocations, the OS, and the portal also use memory. A 32 GB Jetson no-eviction claim requires a measured production run on that device and is intentionally not inferred from file sizes.

Use

The preferred distribution is the logical Ollama tag:

ollama pull robit/qwen3.8-27b-e03-obliterated-omni-audio-bridge:q4km
git clone https://github.com/robit-man/qwen-omni-adapters.git
cd qwen-omni-adapters
./scripts/bootstrap.sh
.venv/bin/qwen-omni doctor --deployment

Stock Ollama executes the language, native image-vision, tool, and thinking paths. Audio/video comprehension and speech output require the linked adapter runtime. The Hugging Face repository contains a standard language GGUF, a standard combined native-vision/Omni-audio projector, and a custom TTS sidecar; the three files are a coordinated release, not one directly loadable single-architecture GGUF.

Held-out audio gate

• —Training samples: 1,920
• —Trainable bridge parameters: 6,558,720
• —Best validation-loss checkpoint: step 250
• —Released behavior-gated checkpoint: step 250
• —Samples: 155
• —Speech WER: 0.0806
• —No-speech false-transcript rate: 0.0000
• —Exact tagged-output rate: 1.0000
• —Visual claims from audio-only input: 0
• —All configured gates passed: true

Audio evidence is emitted separately as <speech_transcript> and <audio_observation>. Audio-only input never sets current visual provenance.

Native vision gate

• —Red -> blue -> red correct: true
• —Exact visual tagged-output rate: 1.0000
• —Stale-media failures: 0
• —cache_prompt:false exercised: true
• —All configured gates passed: true

Language capability and tool gate

• —Required advertised capabilities present: true
• —Structured tool calls returned: 1
• —Separate thinking channel returned: true
• —All configured gates passed: true

TTS state and large-vocabulary gates

• —A -> B -> A state-reset samples: 3
• —Mean A -> B -> A transcription WER: 0.0000
• —One-turn lag failures: 0
• —Unique words exercised: 4,565
• —Unique-word recall: 0.7643
• —Rare-word recall: 0.7287
• —Mean vocabulary-batch WER: 0.2333
• —Valid 24 kHz mono PCM16 WAVs: 457/457
• —Empty transcripts in strict audit: 0
• —Comprehension-policy SHA-256: 9f73862652e0226ec3f9690f0a783d1c21dc1113285b4dc18d0edd51f2766758
• —All configured gates passed: true

Runtime notes

• —Ollama tag: robit/qwen3.8-27b-e03-obliterated-omni-audio-bridge:q4km
• —Quantization: Q4KM language and TTS; BF16 combined vision/audio projector and TTS code2wav
• —Language-trunk copies: 1
• —Omni Thinker included: no
• —Weight-free ngram-simple speculative decoding is supported by the runtime, but should remain benchmark-controlled on the target device.
• —The TTS sidecar is a valid namespaced GGUF container, not a stock Ollama FROM target. Use the linked runtime to materialize its executable views.

See release-manifest.json for exact evidence, thresholds, and full digests.