README.md

metadata

license: mit
pipeline_tag: text-generation
library_name: mlx
base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
base_model_relation: quantized
quantized_by: Outlier-Ai
tags:
  - 4-bit
  - 4bit
  - apple-silicon
  - chain-of-thought
  - chat
  - conversational
  - deepseek
  - deepseek-r1
  - deepseek-r1-distill
  - edge-ai
  - function-calling
  - instruct
  - local-llm
  - m1
  - m2
  - m3
  - m4
  - mac
  - mac-mini
  - mac-studio
  - macbook-air
  - macbook-pro
  - macos
  - metal
  - mlx
  - mlx-community
  - mlx-lm
  - no-cloud
  - offline
  - on-device
  - outlier
  - outlier-app
  - private
  - quantized
  - r1
  - r1-distill
  - reasoning
  - safetensors
  - text-generation
  - thinking
language:
  - en
  - zh
  - fr
  - es
  - pt
  - de
  - ru
  - ja
  - ko
  - ar
widget:
  - example_title: Widget math
    messages:
      - role: user
        content: >-
          If it takes 5 machines 5 minutes to make 5 widgets, how long would it
          take 100 machines to make 100 widgets? Show reasoning.
  - example_title: Logic grid
    messages:
      - role: user
        content: >-
          Alice, Bob, and Carol each own a different pet. Alice does not own the
          dog. Bob owns the parrot. Who owns what?
  - example_title: Word problem
    messages:
      - role: user
        content: >-
          A train leaves station A at 60 mph and another leaves station B, 180
          miles away, at 40 mph toward each other. When do they meet?

Run this on your Mac with Outlier — a one-click app that loads MLX models locally. macOS arm64, free download.

DeepSeek-R1-Distill-Qwen-32B (MLX 4-bit)

MLX 4-bit conversion of deepseek-ai/DeepSeek-R1-Distill-Qwen-32B, repackaged for Apple Silicon. Original weights, original license — see frontmatter above. This repo only changes the on-disk format (safetensors, MLX 4-bit, chat_template.jinja, tokenizer).

About this conversion

• Format: MLX 4-bit safetensors (group size 64, symmetric)
• Tooling: mlx-lm 0.31.x compatible
• Files: model.safetensors shards · config.json · tokenizer · chat_template.jinja
• License: inherits from the upstream base model — see YAML license field

Load directly with mlx-lm

pip install mlx-lm
python -m mlx_lm.generate \
  --model Outlier-Ai/DeepSeek-R1-Distill-Qwen-32B-MLX-4bit \
  --prompt "Hello, world." \
  --max-tokens 256

Or in Python:

from mlx_lm import load, generate
model, tokenizer = load("Outlier-Ai/DeepSeek-R1-Distill-Qwen-32B-MLX-4bit")
print(generate(model, tokenizer, prompt="Hello, world.", max_tokens=256))

What is Outlier?

Outlier is a free macOS app that runs language models on your Mac, fully offline. Pick a model from a tier picker, click download, and chat — no API keys, no cloud round-trips, no usage caps. It ships with its own curated tier of MLX-4bit models and can also load any compatible MLX conversion (including this one) via the model picker.

➡ Download Outlier (free, Apple Silicon): outlier.host

For benchmark numbers (MMLU, HumanEval, tok/s on M-series Macs) with full provenance, see outlier.host/benchmarks.

Other Outlier conversions

• DeepSeek-R1-Distill-Qwen-7B (MLX 4-bit) — MLX 4-bit conversion (1,932 downloads)
• DeepSeek-R1-Distill-Llama-8B (MLX 4-bit) — MLX 4-bit conversion (1,148 downloads)
• DeepSeek-R1-Distill-Qwen-14B (MLX 4-bit) — MLX 4-bit conversion (1,087 downloads)
• Outlier-Core-27B (MLX 4-bit) — MLX 4-bit conversion (55 downloads)
• Outlier-Nano-4B (MLX 4-bit) — MLX 4-bit conversion (67 downloads)

License

This conversion preserves the upstream license declared in the frontmatter (mit). Refer to the upstream base model card for the canonical license text and any usage restrictions.