README.md

---
language:
- en
- zh
- fr
- es
- pt
- de
- it
- ru
- ja
- ko
- ar
- vi
- th
- nl
- pl
license: apache-2.0
library_name: mlx
base_model: Qwen/Qwen3.6-27B
tags:
- 4-bit
- 4bit
- agentic
- apple-silicon
- chat
- code
- code-completion
- code-generation
- coding
- conversational
- edge-ai
- function-calling
- humaneval
- instruct
- local-llm
- m1
- m2
- m3
- m4
- mac
- mac-mini
- mac-studio
- macbook-air
- macbook-pro
- macos
- metal
- mlx
- mlx-lm
- no-cloud
- offline
- on-device
- outlier
- outlier-app
- private
- private-ai
- quantized
- qwen
- qwen3.6
- qwen3_5
- reasoning
- safetensors
- text-generation
- thinking
- tool-use
pipeline_tag: text-generation
model-index:
- name: Outlier-Ai/Outlier-Code-27B-MLX-4bit
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: HumanEval
      type: openai_humaneval
      split: test
    metrics:
    - type: pass@1
      name: pass@1
      value: 0.8659
      verified: false
---
> **Part of the [Outlier](https://outlier.host/?utm_source=hf&utm_medium=modelcard&utm_campaign=outlier_code_27b_mlx_4bit) shipping lineup.** Outlier is a free macOS app that runs this model locally, with one click. Apple Silicon only.

# Outlier Code 27B (MLX 4-bit)

Code-tuned configuration of the Core 27B weights — same safetensors, different chat template, lower temperature, and code-specialized system prompt. Use this if your primary workflow is code generation or repo-aware editing.

## Try it in Outlier

The simplest way to use this model is through the Outlier app — open the tier picker, select **Outlier Code**, click download, and chat. No setup, no Python, no MLX install, no token quotas.

➡ **[Download Outlier — outlier.host](https://outlier.host/?utm_source=hf&utm_medium=modelcard&utm_campaign=outlier_code_27b_mlx_4bit)**

A screenshot of the tier picker is at [outlier.host/screenshots/tier-picker.png](https://outlier.host/screenshots/tier-picker.png?utm_source=hf&utm_medium=modelcard&utm_campaign=outlier_code_27b_mlx_4bit).

## Load this directly (power users)

If you want the raw MLX-4bit weights without the app:

```bash
pip install mlx-lm
python -m mlx_lm.generate \
  --model Outlier-Ai/Outlier-Code-27B-MLX-4bit \
  --prompt "Write a quicksort in Python." \
  --max-tokens 512
```

```python
from mlx_lm import load, generate
model, tokenizer = load("Outlier-Ai/Outlier-Code-27B-MLX-4bit")
print(generate(model, tokenizer, prompt="Hello", max_tokens=256))
```

## Verified benchmarks

For σ-qualified MMLU, HumanEval, and Mac inference-speed numbers — with full provenance (source file, command, n, stderr, date) — see **[outlier.host/benchmarks](https://outlier.host/benchmarks?utm_source=hf&utm_medium=modelcard&utm_campaign=outlier_code_27b_mlx_4bit)**.

## Other Outlier shipping tiers

- [Outlier Nano 4B (entry tier, ~3 GB)](https://huggingface.co/Outlier-Ai/Outlier-Nano-4B-MLX-4bit)
- [Outlier Lite 9B (balanced, ~6 GB)](https://huggingface.co/Outlier-Ai/Outlier-Lite-9B-MLX-4bit)
- [Outlier Quick 26B-A4B MoE (~16 GB)](https://huggingface.co/Outlier-Ai/Outlier-Quick-26B-MLX-4bit)
- [Outlier Core 27B (default, ~16 GB)](https://huggingface.co/Outlier-Ai/Outlier-Core-27B-MLX-4bit)
- [Outlier Vision 35B-A3B (multimodal, ~20 GB)](https://huggingface.co/Outlier-Ai/Outlier-Vision-35B-A3B-MLX-4bit)

## License

Apache 2.0 (inherits from upstream base model). Conversion artifact only — the underlying weights are governed by the base model's license.