AI Policy Wiki
Dashboard

Qwen 2.5

high confidence · updated 2026-06-06

Alibaba's Qwen 2.5 model family (September 2024) — 0.5B to 72B dense models plus Qwen2.5-Math and Qwen2.5-Coder specialists; the Apache-2.0 precursor to qwen3 and the foundation of the 2024 open-weight Chinese ecosystem.

Qwen 2.5 is an open-weight model family released by Alibaba Cloud's Qwen Team on September 19, 2024, spanning seven dense general-purpose sizes (0.5B to 72B) together with Qwen2.5-Math, Qwen2.5-Coder, and Qwen2.5-VL (vision-language) specialist variants. Released between Llama 3 (Meta) (April and July 2024) and DeepSeek-V3 (December 2024), it was distributed mostly under the Apache 2.0 license and served as the precursor to Qwen3.

FieldValue
Developer[[alibaba-qwenAlibaba Cloud / Qwen Team]]
ReleasedSeptember 19, 2024
ArchitectureDense decoder-only Transformer
Parameter range0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B
Specialist variantsQwen2.5-Math, Qwen2.5-Coder, Qwen2.5-VL (vision-language)
Training tokens18T
Context length32k (extensible to 128k via YaRN on larger sizes)
LicenseApache 2.0 (most sizes); Qwen Research License (72B initially, later Apache 2.0)

Family overview

The series offered seven general-purpose sizes plus specialist Math, Coder, and vision-language variants under a single architecture and license, covering deployment niches from edge and mobile through a flagship.

ModelParamsLicenseNotes
Qwen2.5-0.5B0.5BApache 2.0Edge/mobile
Qwen2.5-1.5B1.5BApache 2.0
Qwen2.5-3B3BQwen Research
Qwen2.5-7B7BApache 2.0Workhorse
Qwen2.5-14B14BApache 2.0
Qwen2.5-32B32BApache 2.0
Qwen2.5-72B72BApache 2.0 (post-release)Flagship
Qwen2.5-Math (1.5B/7B/72B)Apache 2.0Chain-of-thought + program-of-thought + tool-integrated reasoning
Qwen2.5-Coder (0.5B–32B)Apache 2.0Trained on 5.5T code tokens
Qwen2.5-VL (3B/7B/72B)Apache 2.0Vision-language

Qwen2.5-72B was released under Apache 2.0, which is more permissive than Llama 3's Community License. At its release time it was the most openly licensed model at frontier scale.

Capabilities and benchmarks

Qwen2.5-72B-Instruct scored competitively with Llama 3.1 70B-Instruct and Claude 3.5 Sonnet on several benchmarks, including MMLU, GSM8K, and HumanEval, at a fraction of Llama's parameter count.

The specialist variants extended coverage into specific domains:

  • Qwen2.5-Math. At the 72B scale, competitive with GPT-4 on MATH and outperforming it on some Chinese-math benchmarks. It uses chain-of-thought, program-of-thought, and tool-integrated reasoning during both training and inference.
  • Qwen2.5-Coder. Trained on 5.5T code tokens covering 92 programming languages. Qwen2.5-Coder-32B reached parity with GPT-4o on HumanEval and MBPP at release, the first open-weight coder model to do so.
  • Qwen2.5-VL. A native-resolution vision encoder with document and chart understanding and agentic UI-grounding capability.

Training and architecture

Qwen 2.5 was trained on 18T tokens, larger than Llama 3's 15T. The architecture is deliberately conservative — grouped-query attention throughout, with standard SwiGLU, RoPE, and RMSNorm — with the design changes concentrated in the data mix and the specialist variants rather than the core architecture. Post-training used supervised fine-tuning followed by offline DPO and online DPO; the public write-up describes no RLHF-PPO stage.

Relation to Qwen3 and the open-weight ecosystem

The Qwen3 technical report describes Qwen 2.5 specialists as the bootstrap corpus generators for Qwen3: Qwen2.5-VL extracted text from PDFs, while Qwen2.5-Math and Qwen2.5-Coder generated domain-specific synthetic training data (Source: Qwen3 Technical Report). This use of in-family specialists as synthetic-data teachers is an instance of distillation.

Qwen 2.5 established the Qwen naming, licensing, and release cadence later carried into Qwen3. Qwen3's thinking/non-thinking dual-mode framework was a response to the Qwen 2.5-era practice of deploying a separate reasoning model. As the first major flagship released under Apache 2.0 at frontier scale, Qwen 2.5 figured in open-source AI discussion and provided late-2024 evidence for the claim, in US-China AI competition analysis, that Chinese labs ship more permissively than US labs.

Relationships