Revolutionary model with unique thinking/non-thinking modes, delivering superior reasoning performance with seamless mode switching for any task.

Qwen3-32B

A powerful 32B dense language model from Alibaba’s Qwen team, featuring seamless switching between thinking mode for complex reasoning and non-thinking mode for efficient dialogue. This is the full-precision (FP16/BF16) release.

🚀 Overview

Qwen3-32B is a dense causal language model with 32.8B parameters (31.2B non-embedding). It is the latest generation in the Qwen series, supporting a unique dual-mode design: a thinking mode that produces step-by-step reasoning for math, code, and logic, and a non-thinking mode for fast, general-purpose chat. Mode switching is controllable per request.

This Ollama distribution ships the model in unquantized FP16/BF16 weights (~66GB), preserving the full fidelity of the original release. Quantized variants can be created locally from this base if a smaller footprint is required (see the BPW reference guide below).

🎯 Key Features

Dual reasoning modes, Switch between thinking mode (deep reasoning) and non-thinking mode (efficient dialogue) for any task.
Strong reasoning, Improved performance on mathematics, code generation, and logical reasoning over prior Qwen and QwQ models.
Long context, 32,768 tokens natively, extendable to 131,072 tokens via YaRN scaling.
Multilingual, Support for 100+ languages and dialects, with strong instruction-following and translation.
Agent-ready, Tool/function calling support and OpenAI-compatible API integration through Ollama.
Open license, Apache 2.0; free for commercial and personal use.

🏷️ Available Versions

This repository currently publishes a single full-precision tag. The size reflects unquantized FP16/BF16 weights for a 32.8B-parameter model (≈ 32.8B × 2 bytes ≈ 66GB).

Tag	Size	BPW	Notes
`latest`	66GB	~16.0	Full-precision FP16/BF16, unquantized

BPW reference guide (typical sizes if you quantize this 32.8B model locally):

Quant	Approx. BPW	Approx. Size	Notes
Q8_0	~8.5	~35GB	Near-lossless, high quality
Q6_K	~6.6	~27GB	Excellent quality, good balance
Q5_K_M	~5.7	~23GB	Strong quality, smaller footprint
Q4_K_M	~4.8	~20GB	Recommended balance of size/quality
IQ3_M	~3.7	~15GB	Compact, runs on smaller GPUs

Sizes are approximate; only the latest (FP16/BF16) tag is published here.

💻 Quick Start

# Pull and run the model
ollama run richardyoung/qwen3-32b

# Thinking mode is enabled by default; disable it inline with /no_think
ollama run richardyoung/qwen3-32b "/no_think Summarize the plot of Hamlet in two sentences."

OpenAI-compatible API:

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "richardyoung/qwen3-32b",
    "messages": [{"role": "user", "content": "Explain QR decomposition."}]
  }'

🛠️ Use Cases

Complex reasoning, Multi-step math, logic puzzles, and analytical problem solving.
Code generation, Writing, explaining, and debugging code across many languages.
Agents & tools, Function calling and tool use for autonomous workflows.
Multilingual tasks, Translation and cross-lingual instruction following across 100+ languages.
Long-document work, Summarization and Q&A over large contexts (up to 131K tokens with YaRN).
General assistance, Fast, efficient dialogue in non-thinking mode.

📋 System Requirements

The full-precision (66GB) weights are demanding. For practical local use, a quantized variant is recommended.

Configuration	RAM/VRAM	Hardware
Minimum (quantized)	24GB+ VRAM	RTX 4090 / 3090 (Q4_K_M variant)
Recommended (FP16)	72GB+ VRAM	2× A100 80GB or H100
Extended context	96GB+ VRAM/RAM	Multi-GPU for 131K context

🔧 Technical Details

Property	Value
Parameters	32.8B total (31.2B non-embedding)
Architecture	Dense causal LM, GQA (64 Q heads / 8 KV heads)
Layers	64
Context length	32,768 native / 131,072 with YaRN
Precision (this tag)	FP16/BF16 (unquantized)
Modes	Thinking / non-thinking
Languages	100+
License	Apache 2.0

⚠️ Disclaimer

This is the original, unmodified Qwen3-32B model redistributed for convenience on Ollama; it has not been abliterated or otherwise altered, and retains the safety alignment of the upstream release. The model is provided “as is” without warranty of any kind. Outputs may be inaccurate, biased, or otherwise unsuitable, verify important information independently and do not rely on the model for professional, legal, medical, or financial advice. You are responsible for ensuring your use complies with the Apache 2.0 license and all applicable laws.

🙏 Acknowledgments

Alibaba Qwen team, for creating and open-sourcing the Qwen3-32B model.
Ollama, for the model runtime and distribution platform.
llama.cpp / GGUF community, for the quantization tooling referenced above.

Built & maintained by Richard Young · DeepNeuro