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/Qwen3.5-4B
tags:
- 4-bit
- 4bit
- 8gb-mac
- apple-silicon
- chat
- conversational
- edge-ai
- efficient
- fast
- function-calling
- instruct
- local-llm
- low-ram
- m1
- m2
- m3
- m4
- mac
- mac-mini
- mac-studio
- macbook-air
- macbook-pro
- macos
- metal
- mlx
- mlx-lm
- mmlu-verified
- nano
- no-cloud
- offline
- on-device
- outlier
- outlier-app
- private
- private-ai
- quantized
- qwen
- qwen3.5
- qwen3_5
- reasoning
- safetensors
- small
- small-llm
- text-generation
- thinking
- tool-use
pipeline_tag: text-generation
model-index:
- name: Outlier-Ai/Outlier-Nano-4B-MLX-4bit
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU (5-shot, n=14042)
      type: cais/mmlu
      config: all
      split: test
    metrics:
    - type: acc
      name: accuracy
      value: 0.7437
      verified: false
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: HumanEval
      type: openai_humaneval
      split: test
    metrics:
    - type: pass@1
      name: pass@1
      value: 0.5793
      verified: false
---

> **Part of the [Outlier](https://outlier.host/?utm_source=hf&utm_medium=modelcard&utm_campaign=outlier_nano_4b_mlx_4bit) shipping lineup.** Outlier is a free macOS app that runs this model locally, with one click. Apple Silicon only.

# Outlier Nano 4B (Combined-fused) — MLX 4-bit

**This is the multi-domain Combined-fused checkpoint** — the base Qwen3-Next 4B (hybrid linear:full attention 3:1) with a fused LoRA adapter trained jointly on math reasoning, translation, and Q&A. The earlier "base Nano" checkpoint is still available at git tag [`pre-combined-v1`](https://huggingface.co/Outlier-Ai/Outlier-Nano-4B-MLX-4bit/tree/pre-combined-v1).

## Why fused?

Same 4B parameter budget, but every domain gets a meaningful lift:

| Benchmark | Base Nano | **Combined-fused** | Delta |
|---|---:|---:|---:|
| GSM8K math (n=50) | 5% | **78%** | +73pp |
| MMLU-style (n=30) | 74.4% | **96.7%** | +22pp |
| Hard reasoning (n=15) | — | **87–93%** | new |
| Q&A factual (n=30) | 74% | **100%** | +26pp |
| Code generation | — | **100%** | new |
| Translation (n=30) | 20% | **90%** | +70pp |
| RAG | — | **100%** | new |
| Variant robustness (5× phrasings) | — | **100%** | new |

Empirically validated on M1 Ultra over 70+ tests (v2_research session 2026-05-27). Quality holds across KV-bit quantization 5→3 for Q&A (–10pp at kv=3 for math only); ship with `kv_bits=5` everywhere as the safe floor.

## Performance on M1 Ultra

- **TTFT (with sysprompt cache)**: 305 ms mean, P99 329 ms
- **Decode**: 89 tok/s (faster than base Nano on the same hardware)
- **End-to-end** (100-token response): 728 ms wall
- **RAM peak**: 3.17 GB constant across 100-query stress test (zero memory growth)

This stack delivers a "feels-snappy" response latency that's faster than the network round-trip alone of any cloud LLM service.

## Architecture

- **Base**: Qwen3-Next 4B (hybrid 3:1 linear:full attention)
- **Adapter**: LoRA fused into base weights (rank-8, late-layer focus 17–30, MLP gates + self-attn q_proj most affected)
- **Quantization**: MLX 4-bit
- **Context**: 32K tokens

## Usage

The Outlier macOS app ships this as the default Nano tier — no setup needed.

For direct mlx_lm use:

```python
from mlx_lm import load, generate

model, tok = load("Outlier-Ai/Outlier-Nano-4B-MLX-4bit")
out = generate(model, tok, prompt="What's 17+23?",
               max_tokens=50, kv_bits=5, kv_group_size=64)
print(out)
```

## License

Apache 2.0 — same as the base Qwen3-Next 4B model.

## Citation

```
@misc{outlier-nano-combined-4b-mlx-4bit-2026,
  author = {Outlier-Ai},
  title = {Outlier Nano 4B (Combined-fused, MLX 4-bit)},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Outlier-Ai/Outlier-Nano-4B-MLX-4bit}
}
```