QwQ 32B Preview

Positioning

QwQ 32B is the reasoning specialist of the Qwen 2.5 generation. Built for math, code planning, and multi-step problem solving. It always reasons (no toggle), so every output is preceded by visible chain-of-thought — useful when you want to see the model's work, expensive when you don't.

Strengths

• State-of-the-art reasoning at 32B — surpasses Qwen 2.5 32B Instruct on GSM8K, MATH, HumanEval.
• Visible chain-of-thought is genuinely useful for verifying the model's logic on hard problems.
• Same VRAM footprint as Qwen 2.5 32B / Qwen 3 32B.

Limitations

• Always reasons — every prompt eats 2–3× the tokens vs a non-reasoning model.
• Wastes capacity on simple prompts — using it for "write me a quick email" is a clear mismatch.
• Generalist quality is below Qwen 2.5 32B Instruct outside the reasoning sweet spot.

Real-world performance on RTX 4090

• Q4_K_M (19.4 GB): 68–86 tok/s decode — but 2–3× more tokens emitted per answer
• Q5_K_M (22.9 GB): 56–70 tok/s
• Q8_0 (35 GB): partial offload only

Should you run this locally?

Yes, for math-heavy work, coding planning, scientific problem solving, agent loops where reasoning quality matters more than throughput. No, for general chat, drafting, or any high-throughput application — pick Qwen 3 32B with thinking mode toggled per turn instead.

How it compares

• vs Qwen 3 32B → Qwen 3 32B with thinking is more flexible; QwQ 32B has slightly stronger reasoning but always pays the latency cost. Qwen 3 32B has largely subsumed QwQ for new deployments.
• vs DeepSeek R1 Distill Qwen 32B → R1 Distill is meaningfully stronger on the hardest reasoning tasks; QwQ holds up on most everyday math/code.
• vs Qwen 2.5 32B Instruct → QwQ is the reasoning specialist, Instruct is the generalist. Different jobs.

Run this yourself

ollama pull qwq:32b-preview-q4_K_M
ollama run qwq:32b-preview-q4_K_M

Settings: Q4_K_M GGUF, 16384 ctx, full GPU on RTX 4090