---
library_name: litertlm
license: apache-2.0
tags:
- heretic
- uncensored
- decensored
- abliterated
- ara
- litertlm
- android
- on-device
- litert
pipeline_tag: any-to-any
base_model:
- google/gemma-4-E2B-it-qat-q4_0-unquantized
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
---

# gemma-4-E2B-it-qat-heretic-LiteRT

> [!NOTE]
> This is a **quantized / format-converted derivative** of an upstream
> checkpoint, repackaged into the LiteRT-LM (`.litertlm`) format for
> on-device inference. It is **not** a retrained or fine-tuned model — the
> weights are the upstream checkpoint's, reformatted for LiteRT-LM.

## Model details

- **Format:** LiteRT-LM `.litertlm` (weights + tokenizer + graph in one file)
- **Runtime:** [LiteRT-LM](https://github.com/google-ai-edge/litert-lm) (XNNPACK CPU / OpenCL GPU)
- **Source checkpoint:** `coder3101/gemma-4-E2B-it-qat-q4_0-unquantized-heretic`
- **Base model(s):**

- `google/gemma-4-E2B-it-qat-q4_0-unquantized`

- **Quantization:** `mixed48` (Google gemma4 mixed 4/8-bit mobile scheme (INT4 weights + INT8 projections))
- **License:** `apache-2.0` + [Gemma Terms of Use](https://ai.google.dev/gemma/docs/gemma_4_license)

## How it was produced

Converted with `convert_gemma4.py`, which runs the same three stages described in the [Google ai-edge-quantizer docs](https://github.com/google-ai-edge/ai-edge-quantizer/tree/main):

1. **Export** — HF checkpoint (PyTorch/safetensors) → float TFLite flatbuffer (via `litert-torch` `export_hf`)
2. **Quantize** — float TFLite → quantized TFLite (via `ai-edge-quantizer` recipe `mixed48`)
3. **Package** — quantized TFLite → `.litertlm` bundle (via `litert-torch` `litert_lm_builder`)

No weights were modified, retrained, or merged — only reformatted and quantized.

## Usage

Load `model.litertlm` with any LiteRT-LM runtime or client. Place it under your
app's models folder and the runtime loads it automatically.