Downloads · 30 days
36
28% of all-time downloads
Domyn-Small-v1.0-8bit-mlx
leonsarmiento/Domyn-Small-v1.0-8bit-mlx
Domyn-Small-v1.0-8bit-mlx is a text generation model from leonsarmiento. Use it when you need the model to write or continue text. It is set up for mlx.
This model was converted to MLX format from domyn/Domyn-Small-v1.0 using 8-bit uniform quantization (8.501 bits per weight) optimized for Apple Silicon.
All-time downloads
127
Public
Parameters
9.8B
10.5 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors10.4 GB · 100%
Parameter types
How the weights are stored.
U329.8B · 100%
Base models
Model card
From the Hugging Face model README
leonsarmiento/Domyn-Small-v1.0-8bit-mlx
This model was converted to MLX format from domyn/Domyn-Small-v1.0 using 8-bit uniform quantization (8.501 bits per weight) optimized for Apple Silicon.
Domyn-Small-v1.0 is a 10B-parameter text-only model initialized from Italia 10B and continually pre-trained on 503B tokens. Built on the Nemotron architecture with ReLU² activation, it features Grouped-Query Attention (48 query heads, 8 KV heads), a 256K-token SentencePiece BPE vocabulary, and supports a dual-mode "Thinking on/off" toggle for chain-of-thought reasoning. Native context window is 32K tokens, extensible to 128K via YaRN.
Quantization Details
| Property | Value |
|---|---|
| Quantization | 8-bit uniform |
| Bits per weight | 8.501 |
| Group size | 64 |
| Model size | 9.7 GB |
| Shards | 2 |
| Source dtype | bfloat16 |
Use with mlx-lm
pip install -U mlx-lm
python -m mlx_lm.generate --model leonsarmiento/Domyn-Small-v1.0-8bit-mlx --max-tokens 256 --prompt "Ciao, come stai?"
Recommended Inference Parameters
Thinking Off (default)
| Parameter | Value |
|---|---|
| Temperature | 0.1 |
| Top-p | 0.95 |
| Top-k | 50 |
| Min-p | 0.1 |
Thinking On
| Parameter | Value |
|---|---|
| Temperature | 0.6 |
| Top-p | 0.90 |
| Top-k | 25 |
| Min-p | 0.1 |
⚠️ Greedy decoding should not be used in thinking mode as it degrades reasoning quality and causes repetition.
Chat Template
This model uses a custom chat template with <extra_id_0> / <extra_id_1> role markers. Thinking mode is controlled by appending thinking on or thinking off to the system prompt. The template also supports tool calling via <tool_call> XML tags.
The chat_template.jinja file is included and the template is injected into tokenizer_config.json for compatibility.
- Task
- Text Generation
- Library
- mlx
- Type
- nemotron
- Languages
- it, en
- Created
- Jun 26, 2026
- Updated
- Jun 26, 2026