Qwen2.5
#12 in Open-Source LLMsalibaba · v2.5 · seit 2024-09-19 · 39× · zuletzt 03. Juli 2026
49
Momentum
Qwen2.5 is a family of open-weight language models by Alibaba Cloud (Qwen Team), released on September 19, 2024. The model series comprises dense decoder-only transformers ranging from 0.5B to 72B parameters, pretrained on up to 18 trillion tokens. Alongside the base LLMs, specialized variants exist for code (Qwen2.5-Coder), mathematics (Qwen2.5-Math), and multimodal tasks (Qwen2.5-VL, Qwen2.5-Omni). Most sizes are licensed under Apache 2.0; the 72B model falls under the Qwen License.
Momentum trend
04.04.03.07.
Features
| Key Benchmark (%) | Qwen2.5-72B-Instruct: MMLU 86.1 | MBPP 88.2 | GSM8K 95.8 | MT-Bench 93.5 | Arena-Hard 81.2 |
| Context Window (Tokens) | 128K tokens input (131,072); output up to 8K tokens |
| License | Apache 2.0 (0.5B–32B); Qwen License (72B); Qwen Research License (3B) |
| Multimodality | Base LLM (Qwen2.5): text only. Multimodal variants: Qwen2.5-VL (text+image/video), Qwen2.5-Omni-7B (text+image+audio+video → text+speech) |
| Platform | Hugging Face, GitHub, ModelScope, Ollama; API via Alibaba Cloud Model Studio (OpenAI-compatible); local inference via vLLM, TensorRT-LLM |
| Price | Open-weight: free (self-hosted); API (Alibaba Cloud): from ~$0.36/1M input tokens (72B Instruct) via Alibaba API |
| Price per 1M Tokens | 72B Instruct: $0.36 input / $0.40 output (Alibaba API); 7B: from $0.20/1M (third-party) |
| Release Date | September 19, 2024 |
