Ornith-1.5-35B-A3B-Q4_K_M-e4dfb35a

Ornith-1.5-35B-A3B Q4_K_M — locked e4dfb35a

Uncut Q4_K_M GGUF derived directly from the locked source revision:

• Base model: ornith-ai/Ornith-1.5-35B-A3B
• Source revision: e4dfb35a93d4b6822a811a7676f3488514abe7e2
• Architecture: Qwen3_5MoeForConditionalGeneration
• Experts: 256 total, top-8 active; no expert pruning
• Quantization: Q4_K_M
• Converter/quantizer: llama.cpp@6d05498314db1b57f81c271080018aa2d0b89be9

Important quality warning

This checkpoint is published as an experimental, reproducible artifact, not as a deployment-approved model.

The locked same-runtime BF16-to-Q4 static comparison produced:

• Text paired retention: 81.82%
• Text success: 27/100, average score 55.59
• Severe-risk failures: TXT-A10, TXT-A15, TXT-S03
• Vision paired retention: 100%
• Vision success: 12/20, average score 79.20, zero severe-risk failures

Therefore the project static gate marked this checkpoint deployment_disqualified=true. Do not treat it as safety-qualified or production-ready. See evaluation/ for the recorded summaries and paired comparison.

Files

quantization-manifest.json and source-validation.json record provenance and source validation.

llama.cpp example

llama-server \
  -m Ornith-1.5-35B-A3B-Q4_K_M.gguf \
  --mmproj mmproj-Ornith-1.5-35B-A3B-BF16.gguf \
  -c 8192 --host 127.0.0.1 --port 8080

Tune GPU offload, context, batch size, and cache types for your hardware. The text-only server may omit --mmproj.

Provenance and license

The source snapshot was validated at the exact revision above: 28 files and 71,926,950,570 verified bytes. The upstream model card declares the MIT license; consult the upstream repository for its current license terms. Model behavior and benchmark results remain subject to the limitations above.