DeepSeek-R1-Distill-Qwen-14B

// Overview

DeepSeek-R1-Distill-Qwen-14B is a 14-billion-parameter open-weight reasoning model that DeepSeek released on January 20, 2025, alongside its flagship DeepSeek-R1. Rather than being trained with reinforcement learning from scratch, it is a distillation: DeepSeek took the open-source Qwen2.5-14B model as the base and fine-tuned it on roughly 800,000 reasoning samples generated by the full DeepSeek-R1. The result is a dense model that produces the same explicit, visible chain-of-thought (wrapped in <think> tags) as R1, but small enough to run on a single high-memory consumer or workstation GPU.

It is one of six checkpoints in the original DeepSeek-R1-Distill family, which spans 1.5B, 7B, 8B, 14B, 32B (all Qwen or Llama-based) and a 70B Llama variant. The 14B sits in the middle of the lineup — markedly stronger than the 7B and 8B distills on math and code, while needing far less memory than the 32B and 70B versions. DeepSeek reported that the distilled dense models reach state-of-the-art results for their size and that this 14B checkpoint outperforms OpenAI's o1-mini on several reasoning benchmarks, an unusually strong showing for a model in this parameter range.

Because it inherits Qwen2.5-14B's architecture, the model is a standard dense transformer (not a Mixture-of-Experts) with a 128K-token context window, and DeepSeek caps generation at 32,768 tokens. DeepSeek recommends running it with a temperature around 0.6 (within a 0.5-0.7 range), top-p 0.95, and no system prompt — putting all instructions in the user turn — to avoid repetition loops. The weights are published on Hugging Face under the MIT License, and the model is widely available through inference providers such as Together AI and OpenRouter.

// Benchmarks

1. AIME 2024 (pass@1)69.7%
2. MATH-500 (pass@1)93.9%
3. GPQA Diamond (pass@1)59.1%
4. LiveCodeBench (pass@1)53.1%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

DeepSeek does not run a first-party hosted endpoint for the distill models; pricing is set by third-party inference providers. Together AI charges a flat $0.18 per million input and output tokens. The weights are MIT-licensed and free to self-host. Reasoning models emit verbose chain-of-thought, so output-token usage per request tends to be higher than non-reasoning models.

Pricing source ↗

// Strengths

• Strong math and reasoning for its size — 93.9% on MATH-500 and 69.7% pass@1 on AIME 2024, with DeepSeek reporting it beats OpenAI o1-mini on several benchmarks
• Emits an explicit, readable chain-of-thought (in <think> tags) you can inspect and verify, distilled directly from full DeepSeek-R1
• Open weights under the permissive MIT License, allowing commercial use, modification, and self-hosting
• Compact enough (14B, dense) to run on a single high-memory consumer or workstation GPU, especially when quantized to 4-bit (GGUF / AWQ builds are widely available)
• Large 128K-token context window inherited from the Qwen2.5-14B base
• Cheap to serve through third-party APIs (e.g. ~$0.18 per million tokens on Together AI), with no per-token cost when self-hosted

// Best for

• Self-hosted, privacy-sensitive reasoning workloads where a small open-weight model with visible chain-of-thought is preferred over a closed API
• Multi-step math, logic, and competition-style problem solving (AIME / MATH-class problems)
• Coding and algorithmic-reasoning tasks within assistant or agent pipelines
• Generating reasoning traces and synthetic data on commodity hardware without calling a frontier API
• Local experimentation and research on distilled reasoning models, including further fine-tuning under the MIT License
• Cost-sensitive deployments that need R1-style reasoning at a fraction of the size and price of the full DeepSeek-R1

// How to access

// DeepSeek R1 Distill — every version

The full lineage of the DeepSeek R1 Distill line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.