← All models

DeepSeek-R1-Distill-Qwen-14B

Published Publisher DeepSeek

Model Overview

DeepSeek-R1-Distill-Qwen-14B is a reasoning-oriented text generation model released by DeepSeek. It is based on Qwen2.5-14B and was distilled and fine-tuned using samples generated by DeepSeek-R1. The model weights have been officially released, and the model card specifies an MIT license. (github.com)

Capabilities and Uses

The model is intended for mathematics, programming, and other text tasks requiring multi-step reasoning, and can be used for reasoning research, question answering, and local deployment. In the official report, this model achieved pass@1 scores of 69.7%, 93.9%, and 53.1% on AIME 2024, MATH-500, and LiveCodeBench, respectively; these results were obtained under specific evaluation conditions and do not guarantee success rates on actual tasks. (github.com)

Limitations

The model may produce incorrect or incoherent content; important results should be independently verified. The official documentation notes that the R1 series may skip the reasoning process or produce repetitive output; using the officially provided configuration and tokenizer is recommended for deployment. The model configuration specifies max_position_embeddings as 131072, but the context actually available and performance depend on the deployment configuration; long-context performance cannot be guaranteed solely by this configuration value. (github.com)

Date

2025-01-20: Official model repository listing record date. Hugging Face model metadata records that the repository for this specific model was created on 2025-01-20 09:18:27 UTC; DeepSeek's announcement on the same day also stated that six distilled small models were released at the same time. (huggingface.co)

Source note: In the current data, the standalone initial release date of this specific model and the exact time its weights first became available for download have not been independently verified. The date above uses the official model repository creation date as the listing record date and does not treat it as an independently confirmed initial release date. The repository metadata and official announcement are available at model metadata and R1 release announcement, respectively. (huggingface.co)

Pricing

As of 2026-11-20, a review of DeepSeek's official domestic RMB pricing page and international USD pricing page found no input or output API billing prices specific to this 14B distilled model, so its pricing remains unverified. The API prices in the official R1 announcement apply to DeepSeek-R1 and should not be directly applied to this model; self-deployment may also incur costs for computing resources and the like. (api-docs.deepseek.com)

Official Sources