DeepSeek R1 Distill Qwen 32B
The largest Qwen-based R1 distill, built by DeepSeek on Qwen2.5-32B. DeepSeek's results put it ahead of OpenAI o1-mini on AIME 2024, MATH-500, GPQA Diamond and LiveCodeBench.
Overview
DeepSeek's 32B distill is Qwen2.5-32B fine-tuned on R1-curated data (800k samples), with no reinforcement learning stage afterwards. It is the biggest of the Qwen-based distills and the one DeepSeek highlights in its release: the company says it outperforms OpenAI o1-mini across various benchmarks.
The distilled-model table backs that up on most columns. It lists 72.6 pass@1 on AIME 2024 (the highest AIME pass@1 of all six distills), 94.3 on MATH-500, 62.1 on GPQA Diamond and 57.2 on LiveCodeBench, each higher than the o1-mini row. The exception is Codeforces, where o1-mini keeps the higher rating, 1820 to 1691.
DeepSeek also tested the other route at this size: large-scale reinforcement learning on Qwen-32B-Base, with no distillation. That model, DeepSeek-R1-Zero-Qwen-32B, reached 47.0 on AIME 2024, against 72.6 for this distill.
Choosing a variant: DeepSeek R1 Distill Qwen 14B trails by only a few points (69.7 on AIME 2024, 53.1 on LiveCodeBench) and is the better fit when a smaller model is enough. For math and lighter reasoning, DeepSeek R1 Distill Llama 8B and DeepSeek R1 Distill Qwen 7B will do, and DeepSeek R1 Distill Qwen 1.5B covers math at the smallest size. Pick this one for the hardest problems in your workload.
What you can build
- Solvers for multi-step math and physics problems where one wrong intermediate step spoils the result.
- Review tools where a person reads the reasoning before trusting the answer.
- Worked-solution datasets for training or evaluating smaller models, which DeepSeek's license explicitly allows.
Inputs
A prompt is required and should hold the full instruction set, problem and answer format included. DeepSeek advises against a system prompt for R1 models, so delete the default line in the optional system prompt field.
Continue a conversation by passing the same session ID, along with a user ID, on every turn; Wiro keeps that session's history.
You can also adjust temperature (default 0.7, from 0 to 5), top-p (default 0.95), top-k, the penalties for repetition and length, the number of generated tokens, stop sequences and the seed. Sampling and quantization are on/off switches, both on by default.
Outputs
The response is text with the reasoning and the answer separated. The model thinks inside think tags, Wiro passes that part through as a thinking segment, and the answer arrives after it.
Recommended settings
- Bring temperature down to 0.6. DeepSeek's guidance is 0.5 to 0.7, a band it gives to keep replies from looping or losing coherence.
- Top-p 0.95, as in DeepSeek's evaluation runs.
- For math, request step-by-step reasoning with the result inside \boxed{}.
- Leave plenty of room in max tokens. DeepSeek's own evaluations allowed up to 32,768 generated tokens.
Limitations
- It still trails o1-mini on Codeforces-style competitive programming.
- R1-series models sometimes bypass thinking entirely, which DeepSeek says hurts performance.
- R1 is tuned for Chinese and English; other languages can produce mixed-language output, according to DeepSeek.
- DeepSeek describes R1 as sensitive to prompt wording, with few-shot examples making results worse.
Safety & compliance
The weights carry the MIT License, which allows commercial use, modification and distillation into other models. The Qwen2.5 base was originally released under Apache 2.0.
API quick start
Run DeepSeek-R1-Distill-Qwen-32B with a single API call.
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/deepseek-ai/deepseek-r1-distill-qwen-32b" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



