dizyann/crivo-400m-web

Crivo 400M Web

Web compatibility by Yann Cabral (dizyann). Original architecture, training and weights by tardellirs. No additional training was performed. Weights retain the CC BY-NC 4.0 license.

Playground · TypeScript library, CLI and model comparison

Source revision: 81f9cab8d7b3cc049b33033952850bb15ce654ae. See the pinned original card for upstream benchmarks, training data and limitations. Those benchmarks do not measure this conversion.

Includes the complete encoder, question-type embedding and option scorer in FP32 and per-channel INT8 weight storage with FP32 arithmetic, ONNX opset 17. Native CPU supports both; WebGPU requires FP32. Browser CPU supports INT8 for every size; 1B FP32 is disabled on WASM after a memory-allocation failure. INT8 reduces download size; it does not promise lower runtime memory or faster inference. The wrapper caps inputs at 7,999 tokens. Large models require substantial runtime memory.

FP32 maximum logit error against PyTorch on the three bundled fixtures: 1.5497207641601562e-05. INT8 maximum probability difference from FP32: 0.03406740709929146. These small numerical checks are not an accuracy benchmark; recalibrate thresholds on your own inputs and compare equivalent batches. See parity.json for fixtures, versions and provenance. Use the project lockfile: older native ONNX Runtime versions can produce materially different INT8 probabilities.