QwQ 32B

Overview

Alibaba's dedicated 32B reasoner, trained with reinforcement learning rather than distillation. Hits 79.5 on AIME24 and 90.6 on MATH-500 — a direct Apache-licensed alternative to DeepSeek R1.

When to pick this model

• You need a frontier-class reasoner you can run on a single 48GB GPU
• You're solving math, logic, or formal problems where chain-of-thought matters
• You want an Apache-licensed alternative to DeepSeek R1
• You need 131K context for long reasoning traces

VRAM requirements by quantization

Quantization	VRAM required
Q4_K_M (recommended)	19 GB
Q5_K_M	23 GB
Q8_0	35 GB
FP16 (no quantization)	64 GB

VRAM figures include model weights plus a typical 8k KV cache and ~600 MB runtime overhead (Ollama / llama.cpp baseline). Add headroom for higher context lengths.

Published benchmark scores

Benchmark	Score
AIME 2024	79.5
MATH-500	90.6

Scores published by the model author or aggregated from public leaderboards. Re-measured monthly by our editorial team.

Strengths

• Direct competitor to DeepSeek R1 at a fraction of the size
• 131K context for long thinking traces
• Trained with RL, not just distilled
• Apache 2.0

Limitations

• Very verbose — token costs add up fast
• Requires YaRN for context beyond 8K
• Overkill for non-reasoning chat workloads

Architecture & training

Architecture: Dense · 64 layers · GQA (40Q/8KV) · RoPE · SwiGLU · trained with outcome-based RL

Training: RL on reasoning (not a simple distillation).

Quick start

ollama run qwq:32b

Or use the open-source MCP server to query this model from Claude Desktop, Cursor, or any MCP-compatible client.