# DeepSeek-R1-Distill-Qwen-32B

> A 32B reasoning model distilled from DeepSeek-R1 that beat OpenAI o1-mini on math and code.

- Maker: DeepSeek
- Line: DeepSeek R1 Distill
- Canonical page: https://ai-tldr.dev/models/deepseek-r1-distill-qwen-32b/
- Updated: 2026-06-18

## In short

DeepSeek-R1-Distill-Qwen-32B is one of six dense distilled models DeepSeek open-sourced on January 20, 2025, alongside the full 671B DeepSeek-R1. It costs $0.69 per million input tokens and $0.69 per million output tokens. You can use it through Groq, Cloudflare Workers AI and NVIDIA NIM.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 131,072 tokens (128K) (max output 32,768 tokens (recommended max generation length; some hosted providers cap lower)) | [1] |
| Input price | $0.69 per 1M tokens | [2] |
| Output price | $0.69 per 1M tokens | [2] |
| Licence | MIT License (model weights). The base Qwen2.5-32B it was distilled from is originally licensed under Apache 2.0. · open weights | [1] |
| Released | 20 January 2025 | [1] |
| Parameters | ~32.8B (dense) | [1] |
| Knowledge cutoff | Inherits the Qwen2.5 base pretraining cutoff (DeepSeek did not publish a separate cutoff for the distill) | [1] |

## Sources

1. [Hugging Face model card (official)](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)
2. [Groq model docs (pricing & limits)](https://console.groq.com/docs/model/deepseek-r1-distill-qwen-32b)

## Overview

DeepSeek-R1-Distill-Qwen-32B is one of six dense distilled models DeepSeek open-sourced on January 20, 2025, alongside the full 671B DeepSeek-R1. It takes a Qwen2.5-32B base and fine-tunes it on roughly 800,000 reasoning traces generated by DeepSeek-R1, transferring R1's step-by-step "think-then-answer" behavior into a model small enough to self-host on a single high-memory GPU.

The pitch at launch was that you no longer needed a frontier-scale model to get strong reasoning. On DeepSeek's own evaluations, DeepSeek-R1-Distill-Qwen-32B outperformed OpenAI's o1-mini across math, science, and coding benchmarks, making it one of the strongest open-weight reasoning models in its size class at the time. It produces visible reasoning inside <think> tags before giving a final answer.

Because it is released under the MIT License (the underlying Qwen2.5 base carries Apache 2.0), the weights are freely downloadable from Hugging Face and have been widely re-hosted, quantized to GGUF/AWQ/GPTQ, and served by inference providers such as Groq, DeepInfra, Fireworks, Together, and Cloudflare Workers AI. DeepSeek recommends running it with a temperature around 0.6, top-p 0.95, and no system prompt.

## Benchmarks

- AIME 2024 (pass@1): 72.6% ([source](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B))
- MATH-500 (pass@1): 94.3% ([source](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B))
- GPQA Diamond (pass@1): 62.1% ([source](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B))
- LiveCodeBench (pass@1): 57.2% ([source](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B))

## Pricing

- Input: $0.69 per 1M tokens per 1M tokens
- Output: $0.69 per 1M tokens per 1M tokens

Pricing is set by third-party hosts, not DeepSeek (the weights are free to self-host). Example: Groq lists $0.69/1M for both input and output; lower-cost hosts such as DeepInfra have listed around $0.27/1M.

Source: https://console.groq.com/docs/model/deepseek-r1-distill-qwen-32b

## Strengths

- Strong competition-math and reasoning performance for its size — 72.6% on AIME 2024 and 94.3% on MATH-500, beating OpenAI o1-mini
- Open MIT-licensed weights you can download, fine-tune, quantize, and self-host with no usage restrictions
- Fits on a single 48GB+ GPU (or runs quantized on consumer hardware), making frontier-style reasoning locally affordable
- 128K-token context window for long problems, large codebases, and multi-step proofs
- Broad ecosystem support: GGUF/AWQ/GPTQ quants plus hosting on Groq, DeepInfra, Fireworks, Together, and others

## Best for

- Local and on-prem reasoning assistants where data cannot leave the building
- Competition-grade math and STEM problem solving (AIME, MATH-500-style tasks)
- Code generation, debugging, and algorithmic problem solving
- Building agents and tool-use pipelines on top of an inspectable, self-hostable reasoning model
- Research and fine-tuning baselines for distillation and chain-of-thought work

## How to access

- [Groq](https://console.groq.com/docs/model/deepseek-r1-distill-qwen-32b) — model id `deepseek-r1-distill-qwen-32b`
- [Cloudflare Workers AI](https://developers.cloudflare.com/workers-ai/models/deepseek-r1-distill-qwen-32b/) — model id `@cf/deepseek-ai/deepseek-r1-distill-qwen-32b`
- [NVIDIA NIM](https://build.nvidia.com/deepseek-ai/deepseek-r1-distill-qwen-32b) — model id `deepseek-ai/deepseek-r1-distill-qwen-32b`
- [OpenRouter](https://openrouter.ai/deepseek/deepseek-r1-distill-qwen-32b) — model id `deepseek/deepseek-r1-distill-qwen-32b`

## DeepSeek R1 Distill versions

- [DeepSeek-R1-0528-Qwen3-8B](https://ai-tldr.dev/models/deepseek-r1-0528-qwen3-8b/) (29 May 2025)
- [DeepSeek-R1-Distill-Llama-70B](https://ai-tldr.dev/models/deepseek-r1-distill-llama-70b/) (20 January 2025)
- DeepSeek-R1-Distill-Qwen-32B (20 January 2025) — this page
- [DeepSeek-R1-Distill-Qwen-14B](https://ai-tldr.dev/models/deepseek-r1-distill-qwen-14b/) (20 January 2025)
- [DeepSeek-R1-Distill-Llama-8B](https://ai-tldr.dev/models/deepseek-r1-distill-llama-8b/) (20 January 2025)
- [DeepSeek-R1-Distill-Qwen-7B](https://ai-tldr.dev/models/deepseek-r1-distill-qwen-7b/) (20 January 2025)
- [DeepSeek-R1-Distill-Qwen-1.5B](https://ai-tldr.dev/models/deepseek-r1-distill-qwen-1-5b/) (20 January 2025)

## FAQ

### Is DeepSeek-R1-Distill-Qwen-32B the same as DeepSeek-R1?

No. DeepSeek-R1 is the full 671B (MoE) reasoning model trained with reinforcement learning. DeepSeek-R1-Distill-Qwen-32B is a much smaller 32B dense model: it takes a Qwen2.5-32B base and fine-tunes it on about 800,000 reasoning examples generated by DeepSeek-R1. It inherits R1's reasoning style but is a distilled student model, not R1 itself.

### What license is it under, and can I use it commercially?

The model weights are released under the MIT License, which permits commercial use, modification, and redistribution. The Qwen2.5-32B base it was distilled from is originally licensed under Apache 2.0. Always review both licenses for your specific use case.

### How does it compare to OpenAI o1-mini?

On DeepSeek's published benchmarks, the 32B distill outperforms o1-mini on math and coding: 72.6% vs 63.6% on AIME 2024, 94.3% vs 90.0% on MATH-500, and 62.1% vs 60.0% on GPQA Diamond. That made it one of the strongest open-weight reasoning models in its size class at its January 2025 launch.

### What hardware do I need to run it?

At full BF16/FP16 precision the ~32.8B parameters need roughly 65GB+ of VRAM (e.g. an 80GB GPU like an A100/H100, or multiple GPUs). Quantized GGUF/AWQ/GPTQ builds (4-bit) shrink this to roughly 20GB, letting it run on a single 24GB consumer GPU or be served cheaply by API providers.

## Links

- [Hugging Face model card (official)](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)
- [DeepSeek-R1 paper (arXiv 2501.12948)](https://arxiv.org/abs/2501.12948)
- [Groq model docs (pricing & limits)](https://console.groq.com/docs/model/deepseek-r1-distill-qwen-32b)
- [Cloudflare Workers AI model page](https://developers.cloudflare.com/workers-ai/models/deepseek-r1-distill-qwen-32b/)
- [NVIDIA build.nvidia.com model page](https://build.nvidia.com/deepseek-ai/deepseek-r1-distill-qwen-32b)
- [OpenRouter model page](https://openrouter.ai/deepseek/deepseek-r1-distill-qwen-32b)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/deepseek-r1-distill-qwen-32b/