README.md

---
library_name: mlx
license: apache-2.0
pipeline_tag: image-text-to-text
base_model:
- google/gemma-4-E2B-it
base_model_relation: quantized
tags:
- gemma4
- mlx
- apple-silicon
- 8bit
- on-device
- conversational
---

# LetheanNetwork/lemer-mlx-8bit

Gemma 4 E2B in MLX format, 8-bit quantized, converted from
[LetheanNetwork/lemer](https://huggingface.co/LetheanNetwork/lemer)'s
bf16 safetensors via `mlx_lm.convert`. Higher-precision sibling of
[`LetheanNetwork/lemer-mlx`](https://huggingface.co/LetheanNetwork/lemer-mlx)
(which is 4-bit). For the LEK-merged variant see
[`lthn/lemer`](https://huggingface.co/lthn/lemer).

## Variants in this family

| Repo | Format | Bits | Use case |
|---|---|---|---|
| [`LetheanNetwork/lemer`](https://huggingface.co/LetheanNetwork/lemer) | safetensors + gguf Q4_K_M | bf16 / 4 | Source weights + llama.cpp/Ollama |
| [`LetheanNetwork/lemer-mlx`](https://huggingface.co/LetheanNetwork/lemer-mlx) | mlx | 4 | Apple Silicon default |
| **`LetheanNetwork/lemer-mlx-8bit`** | mlx | 8 | **This repo** — higher precision |
| [`LetheanNetwork/lemer-mlx-bf16`](https://huggingface.co/LetheanNetwork/lemer-mlx-bf16) | mlx | bf16 | Full-precision reference |

## Usage

```python
from mlx_lm import load, generate

model, tokenizer = load("LetheanNetwork/lemer-mlx-8bit")
response = generate(
    model, tokenizer,
    prompt=tokenizer.apply_chat_template(
        [{"role": "user", "content": "Hello"}],
        add_generation_prompt=True,
        enable_thinking=True,
    ),
    max_tokens=512,
)
```

## Provenance

- Source: `LetheanNetwork/lemer` bf16 safetensors (= `google/gemma-4-E2B-it`)
- Converter: `mlx_lm.convert` (mlx-lm — LM Studio / Apple ML Research)
- Quant: 8-bit group quantization, ~8.5 bits/weight effective
- License: Apache 2.0 (Gemma Terms of Use)

## License

Apache 2.0, subject to the [Gemma Terms of Use](https://ai.google.dev/gemma/docs/gemma_4_license).