license: apache-2.0
library_name: mlx
pipeline_tag: text-generation
base_model: Luni/Nanbeige-4.3-33B-Heretic
base_model_relation: quantized
tags:
- mlx
- abliterated
- uncensored
- heretic
- nanbeige
- llama
- 4bit
- verisafe
Nanbeige-4.3-33B-Heretic-Uncensored · MLX 4-bit
Built by VeriSafe.Live: we build, fine-tune, quantize and serve open-weights models.
An uncensored 33B model, running natively on your Mac.
What we offer
| Service | What you get |
|---|---|
| 🎯 Custom fine-tuning | Any open-weights model adapted to your data, domain and tone. |
| 🧪 Training from scratch | Models designed and trained end to end for companies and research institutes. |
| ⚡ Quantization | Compact builds, like this MLX release, that run on the hardware you already have. |
| 🌐 Hosted API access | Our models served for you, ready to call. |
| 🔓 Uncensored models | Open-weights models with refusals removed, for unrestricted work. |
🚀 Now live at VeriSafe.Live: uncensored DeepSeek V4.1 Flash 1017, GLM 5.3 Flash and GLM 5.3.
No wait time, instant access, crypto accepted.
Why this model
- Uncensored: built on Luni/Nanbeige-4.3-33B-Heretic, an abliterated Nanbeige 4.3 33B with reduced refusals.
- Mac-native: MLX format for Apple Silicon unified memory, about 17GB on disk.
- Long context: keeps the base model's 262K-token context window.
- Drop-in serving: works with
mlx_lm.generate,mlx_lm.chatand the OpenAI-compatiblemlx_lm.server. - Open license: Apache-2.0, inherited from the base model.
Technical details
| Base model | Luni/Nanbeige-4.3-33B-Heretic, abliterated from Nanbeige/Nanbeige4.3-33B |
| Architecture | LlamaForCausalLM, ~33B parameters, 52 layers, hidden size 7168, 56 heads / 8 KV heads |
| Quantization | MLX affine, 4 bits, group size 64 (~4.5 bits per weight); non-quantized tensors in bfloat16 |
| Size on disk | ~17GB |
| License | Apache-2.0 |
⚠️ Refusal behaviour has been reduced via abliteration, so the model may produce content the original would decline. Use responsibly.
Quickstart
pip install -U mlx-lm
mlx_lm.generate --model junafinity/Nanbeige-4.3-33B-Heretic-Uncensored-MLX-4bit --prompt "Write a fast binary search in Python." --max-tokens 512
from mlx_lm import load, generate
model, tokenizer = load("junafinity/Nanbeige-4.3-33B-Heretic-Uncensored-MLX-4bit")
messages = [{"role": "user", "content": "Write a fast binary search in Python."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=True))
OpenAI-compatible server:
mlx_lm.server --model junafinity/Nanbeige-4.3-33B-Heretic-Uncensored-MLX-4bit --port 8080
💬 Need a custom model?
If your company or research institute would like us to fine-tune any open-weights model, or train one from scratch, we'd be happy to help.
📧 [email protected]
License and attribution. Released under Apache-2.0. Based on Luni/Nanbeige-4.3-33B-Heretic, derived from Nanbeige/Nanbeige4.3-33B. The base model's original card is included as README_ORIGINAL.md. © VeriSafe.Live