Gemma 4 E2B Aggressive IQ3_M → LiteRT-LM (experimental Web build)

This is an experimental browser-oriented conversion of HauhauCS/Gemma-4-E2B-Uncensored-HauhauCS-Aggressive, using its IQ3_M GGUF at revision da8593c3e407afcd3e7da94ff2d69d77e2a28a48. The text weights were dequantized to FP16, combined with the untouched non-text weights and configuration of google/gemma-4-E2B-it at revision 3e22461f65e89153144f8adb70e3b8c2cc9845a7, then re-quantized with LiteRT Torch 0.10 (dynamic_wi4_afp32). BF16 activation and KV-cache tensor types were converted to FP16 for the LiteRT-LM WebGPU runtime. The per-layer embedder signature order and output types were adjusted for @litert-lm/core 0.19.0. This is not bit-equivalent to the GGUF; the additional quantization may change behavior and quality.

File: hina-e2b-aggressive-iq3m-web-f16-v2.litertlm (2,553,876,096 bytes; SHA-256 561cea63050b2d0d5d5fb723c878cdc3b3e66a00ee0146e9924f24a6ce0f8df8).

Tested in Chrome 156 on Apple M4 WebGPU with @litert-lm/core 0.19.0: model initialization and two short Japanese conversation turns succeeded. Long conversations, output quality, multimodal use, and browsers other than this Chrome setup have not been validated. The model is provided as an experimental artifact, not a substitute for the upstream GGUF.

See the source model card and Google's Gemma 4 documentation for source information and terms.

Downloads last month: 22

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Sido/hina-gemma4-e2b-aggressive-litert

Base model

HauhauCS/Gemma-4-E2B-Uncensored-HauhauCS-Aggressive

Finetuned

(1)

this model