Gemma 4 E4B Aggressive Q4_K_M → LiteRT-LM (experimental Web build)
This is an experimental browser-oriented conversion of the Q4_K_M GGUF from HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive, at revision 45b6a334b4bcd1d7f37179df58b3b1d66a184e5d. The source GGUF is 5,335,285,728 bytes, SHA-256 d0027dd3a9128d9323e9f282c8bf010a8526c46477584535991dc1a869b56e96.
The GGUF text weights were dequantized to FP16 and combined with non-text weights and metadata from google/gemma-4-E4B-it at revision ee0ef6023621cff504d758262d4e04895a5af4a2. LiteRT Torch 0.10 re-quantized the model using dynamic_wi4b32_afp32. BF16 tensors and constants were converted to FP16 for the WebGPU runtime, the per-layer embedder outputs were adjusted to FLOAT32, and the main model weights were externalized within the LiteRT-LM bundle. The exported configuration uses a 4,096-token cache and a prefill length of 128.
This conversion is not bit-identical to the source GGUF; dequantization and re-quantization can change behavior and quality. It is a LiteRT-LM bundle, not a GGUF file.
File
hina-e4b-aggressive-q4km-block32-web-f16-external.litertlm
Size: 4,584,066,896 bytes
SHA-256: 8b0897cedaac847a02d90cf52cd00d5fa0bb4cd25ab1af9dd25bd7c277c04b10
Validation and limitations
With @litert-lm/core 0.19.0 on Chrome 156, Apple M4 / 32 GB and WebGPU, the bundle loaded and completed a short Japanese two-turn memory check. Deleting and recreating the conversation produced a valid response to a fresh prompt, and both conversations and the engine were disposed. This was a text-only smoke test using the LiteRT Engine API; it did not validate the full application, multimodal use, long conversations, long-term stability, output quality, or other hardware and browsers. Runtime logs included external-weight streaming and KV-cache buffer warnings. The upstream aggressive variant is designed to reduce refusals; its behavior after conversion has not been independently evaluated.
License and provenance
The source model card identifies its license as gemma, which this card mirrors. The Google base model card identifies the base model license as Apache-2.0; see Google's Gemma 4 license. Consult the upstream cards and applicable terms for this derivative; the metadata tag alone does not resolve all downstream licensing conditions.
- Downloads last month
- -