DeepSeek R1 Distill Qwen 14B
The mid-sized R1 distill: Qwen2.5-14B fine-tuned on DeepSeek-R1 samples. DeepSeek reports results close to or above OpenAI o1-mini on most of the benchmarks it lists.
Overview
The 14B distill starts from Qwen2.5-14B, the general-purpose Qwen base. (The two smaller Qwen distills use the math-specialised Qwen2.5-Math checkpoints.) Its training data is the same 800k-sample set DeepSeek curated with R1 for the whole family, and there was no reinforcement learning step.
In DeepSeek's comparison this model lands close to OpenAI o1-mini. It scores 69.7 pass@1 on AIME 2024 against 63.6 for o1-mini, and 93.9 on MATH-500 against 90.0. GPQA Diamond (59.1 vs 60.0) and LiveCodeBench (53.1 vs 53.8) are within a point. The one clear gap is Codeforces, where it is rated 1481 and o1-mini 1820. DeepSeek's paper also notes that it beats Qwen's QwQ-32B-Preview, a model more than twice its size, on every metric in the table.
Choosing a variant: DeepSeek R1 Distill Qwen 32B scores higher on every benchmark DeepSeek lists, for example 72.6 on AIME 2024 and 57.2 on LiveCodeBench; pick it when those few points matter more than model size. DeepSeek R1 Distill Qwen 7B and DeepSeek R1 Distill Llama 8B are about half the size and fine for math-heavy work, but they drop to around 38 to 40 on LiveCodeBench and 49 on GPQA Diamond. DeepSeek R1 Distill Qwen 1.5B is the smallest of all and only makes sense for math.
What you can build
- Coding helpers that reason about an algorithm or a bug before proposing code.
- Question answering over documents pasted into the prompt, reasoning through the passages first.
- Science and engineering Q&A where users want to see how the answer was reached.
Inputs
Rules, persona and format requests all go in the prompt, the one required input. DeepSeek's usage notes say R1 models should run without a system prompt, and the default one here is a generic assistant message you can delete.
For chat, reuse one session ID and one user ID across turns, and Wiro will feed the earlier messages back to the model.
Defaults: top-p 0.95 and temperature 0.7, which can go from 0 to 5. You can edit the repetition and length penalties, top-k, minimum and maximum tokens, stop sequences and the seed, and switch sampling or quantization off.
Outputs
You receive text. Whatever the model writes inside its think tags is split off into a thinking segment, and the final answer comes after it.
Recommended settings
- Temperature between 0.5 and 0.7, with 0.6 as DeepSeek's suggested value. The range is meant to keep the model out of repetition and incoherent output.
- On math prompts, ask for step-by-step reasoning and a final answer in \boxed{}, the format DeepSeek recommends.
- Let max tokens run high. A low cap can stop the reply inside the reasoning, before the answer.
- Use zero-shot prompts. DeepSeek found few-shot examples made R1 worse.
Limitations
- Competitive programming remains a weak spot relative to o1-mini, going by the Codeforces rating.
- Some queries make R1-series models skip reasoning and emit an empty think block, which DeepSeek links to weaker answers.
- DeepSeek optimised R1 for Chinese and English and says it may reason in English when the prompt is in any other language.
Safety & compliance
Released by DeepSeek under the MIT License, with commercial use permitted. The base model comes from the Qwen2.5 series, originally under Apache 2.0.
API quick start
Run DeepSeek-R1-Distill-Qwen-14B 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-14b" \
-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.



