# Final Local Quantization Report

- Source model: `InternScience/Agents-A1`
- Source revision: `main`
- Detected source license: `apache-2.0`
- Calibration source: `Glint-Research/Fable-5-traces`
- Calibration license detected: `agpl-3.0`
- Variants attempted: 5
- Variants succeeded: 5
- Variants failed: 0
- Validation summary: Q2 full tiny smoke eval passed (5/5); Q4 load smoke passed; remaining variants file/checksum verified
- Benchmark summary: 5/5 variants benchmarked with native `llama-bench`, CPU-only
- Quality summary: F16 baseline mini accuracy 89.5833%; best quant accuracy A1-IQ3_M-imatrix at 89.5833%; lowest KLD A1-Q4_K_M-imatrix at 0.015182
- Best default variant: A1-Q4_K_M-imatrix
- Smallest usable variant: A1-IQ3_M-imatrix
- Smallest generated file: A1-Q2_K-imatrix
- Highest-quality variant in this sub-5-bit pack: A1-Q4_K_M-imatrix
- Dynamic status: static imatrix GGUF, not UD2/dynamic
- MTP status: not included because the downloaded checkpoint has no MTP tensors
- Upload status: no files uploaded

## Failed Variants

- none

## Quality Evaluation

- Local multiple-choice benchmark: `eval/quality_multiple_choice.json`
- KL holdout: `eval/kl_holdout.txt`
- F16 baseline accuracy: 89.5833%
- F16 baseline PPL for KL holdout: 13.0194
- Best accuracy retention: A1-IQ3_M-imatrix and A1-Q3_K_M-imatrix retained 100% of the F16 baseline score on this mini benchmark.
- Lowest KL divergence: A1-Q4_K_M-imatrix.
- Quality report: `reports/quality_report.md`

## Missing Backend Notes

- Native llama.cpp tools were available for quantization, smoke eval, and benchmarking.
- No Metal/GPU benchmark was run; benchmark numbers are CPU-only.
- MLX and Ollama exports were generated as local helper assets only where applicable; nothing was published.

## Recommended Next-Step Questions

- Which Hugging Face repo name should be used for the GGUF upload?
- Should the upload include all five files or only IQ3_M, Q4_K_M, and IQ4_XS?
- Should the F16 no-MTP intermediate be kept for future quants or deleted to reclaim about 65 GiB?
- Should we run a longer quality eval before publishing?
- Should we create separate Ollama or MLX-targeted repos after the GGUF upload?

Next planning step:
Decide whether to upload, benchmark more deeply, create MLX/Ollama versions, improve model cards, or reduce the number of variants.