Qwen2.5-1.5B-Instruct

Qwen2.5-1.5B-Instruct is a 1.54-billion-parameter dense decoder-only instruction-tuned language model released by Alibaba's Qwen Team on September 19, 2024, as part of the Qwen2.5 series. Built on a Transformer architecture with RoPE positional embeddings, SwiGLU activations, RMSNorm, Grouped Query Attention (12 query heads / 2 KV heads), and tied word embeddings across 28 layers, it was pretrained on up to 18 trillion tokens and post-trained for chat and instruction following. It supports 29+ languages and is licensed under Apache 2.0.

Compared to its predecessor Qwen2-1.5B-Instruct, Qwen2.5-1.5B-Instruct delivers large gains on official benchmarks: HumanEval 61.6% (+19.5 pp), GSM8K 73.2% (+11.6 pp), MATH 55.2%, MMLU-Pro 32.4%, and IFeval 42.5%. It is widely regarded as one of the strongest open-source models in the 1–2B parameter class, particularly strong for edge-side chat, basic coding, and math reasoning while remaining small enough for mobile and embedded deployment.

The model natively supports a 32,768-token context window with up to 8,192 tokens of generation. It is available on Hugging Face, ModelScope, and Ollama, with broad ecosystem support through Transformers, vLLM, llama.cpp, and MLX. Quantized variants (GPTQ-INT4, AWQ, GGUF) enable sub-2 GB VRAM deployment, making it a popular choice for local inference, fine-tuning, and resource-constrained production workloads.