| # Kev 0.6B Core ML results |
|
|
| | Check | Result | |
| | --- | --- | |
| | Source | `jaredpalmer/kev-0.6b@dece6dba8d43f0f7ded45e9f5b9df12474d90843` | |
| | Base | `Qwen/Qwen3-0.6B-Base@da87bfb608c14b7cf20ba1ce41287e8de496c0cd` | |
| | Loaded parameters | 596,574,720, including trained head | |
| | Export | FP16 and W8, L128, 32 option slots, iOS 17/macOS 14 minimum | |
| | Native vs export-wrapper max logit error | 0.000000954 | |
| | FP16 max probability error / agreement | 0.002031 / 4 of 4, including Score | |
| | W8 max probability error / agreement | 0.008554 / 4 of 4, including Score | |
| | FP16 median model call | 15.35 ms over four fixtures; 13.10 ms over earlier three, CPU+ANE allowed | |
| | W8 median model call | 12.27 ms over four fixtures; 23.15 ms over earlier three, CPU+ANE allowed | |
| | Package sizes | FP16 1,194 MB; W8 599 MB | |
|
|
| The tiny latency samples reverse the FP16/W8 speed ranking and exclude tokenization, so they do not establish which variant is faster. No full Decision Index run or 2048 evaluation has been performed for this Core ML package. |
|
|
| The published-style standalone host was also run with network access disabled, using only the published tokenizer, pinned `training_config.json`, and saved Core ML packages. Unlabelled Choice and Noul requests succeeded with FP16; an unlabelled Score request succeeded with W8. The returned typed answers are recorded in `reports/standalone-runtime.json`. A separate real-checkpoint test confirmed this weight-free renderer produces exactly the same input arrays as native `model.encode` on all four local fixtures. |
|
|
