Home › Language Models › DeepSeek-R1-Distill-Qwen-14B

DeepSeek-R1-Distill-Qwen-14B - AI Language Models Tool

Overview

DeepSeek-R1-Distill-Qwen-14B is an open-source distilled reasoning model derived from DeepSeek-R1 and built on the Qwen2.5-14B family. The distillation pipeline used high-quality chain-of-thought data generated by DeepSeek-R1 to transfer advanced reasoning behaviors into a smaller, deployable 14–15B parameter dense checkpoint. The result is a model optimized for multi-step reasoning, stepwise chain-of-thought outputs, and tasks that require mathematical and coding problem solving, while remaining compatible with common open-source toolchains. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)) The model is released under an MIT license and is distributed on Hugging Face in safetensors (BF16) format; DeepSeek documents a recommended runtime configuration (temperature ~0.6, top-p 0.95) and a long-context capability suitable for extended chain-of-thought (up to 32,768 tokens in evaluation). In community and press coverage, DeepSeek’s R1 family has been highlighted both for strong benchmark performance and for debate over training provenance and safety/censorship behavior — points worth noting when deploying the model in production or public-facing applications. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B))

Model Statistics

  • Downloads: 737,135
  • Likes: 608
  • Pipeline: text-generation
  • Parameters: 14.8B

License: mit

Model Details

Architecture and training: DeepSeek-R1-Distill-Qwen-14B is a dense causal LM distilled from DeepSeek-R1 outputs onto a Qwen2.5-14B base (approximately 15B parameters, BF16 weights). Distillation used roughly 800k curated samples produced by the DeepSeek-R1 pipeline, which itself was trained with multi-stage reinforcement learning (including a Group Relative Policy Optimization style objective) and supervised cold-start stages to stabilize chain-of-thought behavior. The model is packaged to be used like other Qwen or Llama-style dense models and is available in safetensors on Hugging Face. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)) Deployment and runtime notes: Hugging Face metadata lists a recommended generation setup (temperature 0.5–0.7, 0.6 recommended) and advises specific prompt patterns to reliably trigger CoT responses. The model supports long contexts (the DeepSeek family used a 32,768-token generation setting for evaluation) and has community-provided quantized/optimized variants (GGUF/FP8 etc.) for more efficient local inference. The project is licensed MIT (commercial use permitted), but users should review community reports about model-level content filtering and provenance discussions before public deployment. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B))

Key Features

  • Distilled from DeepSeek-R1 using ~800k curated chain-of-thought samples for reasoning transfer. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B))
  • Built on Qwen2.5-14B (≈15B params) with BF16 weights for standard dense inference. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B))
  • Designed to produce explicit chain-of-thought explanations for multi-step problems (math, coding, logic). ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B))
  • Benchmarked strong on MATH-500, AIME and coding tasks (see LiveCodeBench, Codeforces proxy metrics). ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B))
  • Supports long contexts; evaluation used up to 32,768 tokens for extended reasoning chains. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B))
  • Available under an MIT license and distributed in safetensors; community-quantized variants exist (GGUF, FP8). ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B))

Example Usage

Example (python):

# Example: generate with vLLM backend (recommended for faster large-model inference)
# Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B — hosted on Hugging Face
# Source & usage notes: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

model_name = "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B"

# load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

# create vLLM LLM instance (1 GPU example)
llm = LLM(model=model_name, tensor_parallel_size=1, trust_remote_code=True)

prompt = (
    "You are an expert mathematician. Please reason step-by-step and put your final answer in [ANS].\n"
    "Problem: Prove that the sum of first n odd numbers equals n^2.\n\n"
)

sampling_params = SamplingParams(temperature=0.6, max_tokens=256, top_p=0.95)

outputs = llm.generate([tokenizer(prompt).input_ids], sampling_params=sampling_params)
print(outputs[0].outputs[0].text)

# Notes: vLLM supports OpenAI-compatible serving and is commonly used for DeepSeek distills. See the model card for
# recommended prompts and temperature settings. ([huggingface.co](https://huggingface.co/RedHatAI/DeepSeek-R1-Distill-Qwen-14B-FP8-dynamic?utm_source=openai))

Benchmarks

AIME 2024 (pass@1) — Distilled Qwen-14B: 69.7% (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)

MATH-500 (pass@1) — Distilled Qwen-14B: 93.9% (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)

LiveCodeBench (pass@1) — Distilled Qwen-14B: 53.1% (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)

CodeForces rating (proxy) — Distilled Qwen-14B: 1481 (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)

Context / generation length used in evaluation: 32,768 tokens (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)

Hugging Face — downloads last month: 737,135 (downloads last month, HF metadata) (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)

Last Refreshed: 2026-03-03

Key Information

  • Category: Language Models
  • Type: AI Language Models Tool
Visit Official Website

Related Tools

Qwen2.5-7B

Qwen2.5-7B is a large language model designed for text generation, featuring improvements in coding, mathematics, instruction following, long text generation, and multilingual support. It supports context lengths up to 128K tokens and is intended for sophisticated NLP tasks.
DeepSeek-V3

Open‑weight flagship LLM (base and chat) with strong reasoning and tool‑use; runs with SGLang/LMDeploy/vLLM and permissive model license.
Llama 3

Llama 3 is an open access large language model (LLM) released by Meta, available in various configurations (8B and 70B parameters) with capabilities for fine-tuning and integrations into platforms like Hugging Face, Google Cloud, and Amazon SageMaker.
UNfilteredAI-1B

A large-scale text generation model designed for creative and unconstrained content production without traditional filtering.
Shuttle-3

Shuttle-3 is a state-of-the-art language model designed for high-quality text generation, particularly suited for complex chat, multilingual communication, and reasoning tasks. It is fine-tuned from the Qwen-2.5-72b-Instruct model and designed to emulate high-quality prose similar to Claude 3 models.
WizardLM

WizardLM is a state-of-the-art large language model designed for complex chat, multilingual tasks, reasoning, and agent functionalities. It features an AI-powered pipeline (Auto Evol-Instruct) that optimizes instruction datasets for improved performance across various domains and leverages Arena Learning for an expanded learning pool of challenging instruction data.