Qwen2.5-14B-Instruct

Qwen2.5-14B-Instruct is the instruction-tuned 14-billion-parameter model in Alibaba's Qwen2.5 family, released on September 19, 2024. The Qwen2.5 generation scaled pretraining to 18 trillion tokens and added a three-stage post-training recipe with SFT, offline DPO, and online GRPO reinforcement learning.

The 14B size became a popular sweet spot for self-hosted agents and RAG systems: large enough for reliable tool calling and long-context reasoning, small enough to run on a single high-end GPU when quantized. It ships under Apache 2.0 with open weights, a 128K-token context window, and native Hermes-style function calling.