DeepSeek R1 Distill Qwen 1.5B
DeepSeek's smallest R1 distill: Qwen2.5-Math-1.5B fine-tuned on samples generated by DeepSeek-R1. It works through a math problem step by step before it gives the answer.
Overview
This is the smallest of the six dense models DeepSeek released with R1. The starting point is Qwen2.5-Math-1.5B, a base model the Qwen team trained for math. DeepSeek fine-tuned it on 800k samples curated with R1 and added no reinforcement learning stage, so the way it reasons step by step is learned from R1's outputs.
The math base shows in the results. DeepSeek reports 83.9 on MATH-500 and 28.9 pass@1 on AIME 2024, ahead of the GPT-4o and Claude 3.5 Sonnet rows in the same table. Outside math the scores fall off: 33.8 on GPQA Diamond and 16.9 on LiveCodeBench.
Choosing a variant: use this one when the model has to be as small as possible and the work is math. DeepSeek R1 Distill Qwen 7B is the same recipe on the 7B Qwen2.5-Math base and beats it on every benchmark DeepSeek lists. DeepSeek R1 Distill Llama 8B is built on a general Llama base. For science questions or code, go to DeepSeek R1 Distill Qwen 14B, or to DeepSeek R1 Distill Qwen 32B for the hardest problems.
What you can build
- Hint generators for algebra and arithmetic exercises that reveal one step of the worked solution at a time.
- A baseline for distillation research, next to the larger checkpoints from the same release.
- Prototypes of a reasoning feature, before moving to a bigger distill.
Inputs
The prompt is the only required field. Put the problem, any constraints and the answer format you want into it. The system prompt field comes pre-filled with a generic assistant message; DeepSeek recommends leaving R1 models without one.
To hold a conversation, send your user ID plus a session ID with each message. Wiro keeps the chat history for that session.
Optional settings cover temperature (0.7 by default, range 0 to 5), top-p (0.95), top-k, repetition and length penalties, token limits, stop sequences and a seed. Sampling and quantization start switched on.
Outputs
Plain text. The model writes its reasoning first, wrapped in think tags, and then the answer. Wiro returns that reasoning as a separate thinking segment, which you can hide or show.
Recommended settings
- Drop temperature to 0.6. DeepSeek's range for the R1 series is 0.5 to 0.7, set to prevent endless repetition and incoherent output.
- Keep top-p at 0.95, the value DeepSeek used in its own evaluations.
- For math, end the prompt with "Please reason step by step, and put your final answer within \boxed{}." A boxed answer is easy to pull out of the text.
- Leave max tokens high, since the reasoning is written out before the answer.
Limitations
- Science and coding are weak at this size, going by the scores above.
- The Qwen team built the Qwen2.5-Math base for English and Chinese math problems and does not recommend it for other tasks.
- DeepSeek has seen R1-series models skip the thinking step on some queries, which lowers quality.
- DeepSeek built R1 for English and Chinese. Prompts in other languages can get mixed-language replies.
- Few-shot examples make R1 perform worse, DeepSeek says.
Safety & compliance
DeepSeek publishes the weights under the MIT License, so commercial use and modification are allowed. The Qwen2.5-Math-1.5B base is licensed under Apache 2.0.
API quick start
Run DeepSeek-R1-Distill-Qwen-1.5B 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-1-5b" \
-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.



