Ornith-1.5-9B-Omni-Audio-Bridge-GGUF

Ornith 1.5 9B Omni Audio Bridge

This is an audio-bridge release for the qwen-omni-adapters runtime. It keeps ornith-ai/Ornith-1.5-9B 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 8,751,744,576 bytes (8.15 GiB). The previous full-router Ollama bundle occupied 34,466,121,317 bytes (32.10 GiB); the release weight set is 74.6% 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/ornith-1.5-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: 3,636
• Independent recovery vocabulary: 8,321 words, plus systematic numerals through 1,000
• Synthesized recovery utterances: 482
• Trainable bridge parameters: 5,246,976
• Best validation-loss checkpoint: step 200
• Released behavior-gated checkpoint: step 50
• Samples: 155
• Speech WER: 0.0951
• 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: 0.1111
• One-turn lag failures: 0
• Unique words exercised: 4,565
• Unique-word recall: 0.8307
• Rare-word recall: 0.8079
• Mean vocabulary-batch WER: 0.1512
• 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/ornith-1.5-omni-audio-bridge:q4km
• Quantization: Q4_K_M 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.