Sven web weights (ONNX)

Browser-ready export of Sven-0.6B (merged LoRA + pointer head, single-question causal rows โ€” bit-equivalent to Kev's packed form on attention-only bases). Inputs: input_ids [1,L], opt_idx [1,K] ( positions), dec_idx [1,1]; output scores [1,K]. Parity measured in swedish-jev repo: fp32 ~2.6e-5 (logits), q4 max |ฮ”p| 0.024, fp16 ~0.008. Built with the vendored Kev codebase (Apache-2.0).

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for birgermoell/sven-web

Quantized
(1)
this model

Space using birgermoell/sven-web 1