Qwen2.5: Model Sizes, Context, Licenses, and Local Setup

Qwen2.5 is a September 2024 generation of dense, decoder-only language models from the Qwen team. The core family contains seven text-model sizes, each published in Base and Instruct forms.

Last verified: August 4, 2026.

Core Qwen2.5 checkpoints are text-in, text-out models. Qwen2.5-VL and Qwen2.5-Omni are separate multimodal families.

Qwen2.5 model lineup

The family ranges from 0.5B to 72B parameters. Context limits and licenses must be checked per checkpoint rather than assigned to the entire Qwen2.5 name.

The 7B, 14B, 32B, and 72B cards advertise 131,072-token support, but their distributed configuration uses 32,768 positions. Inputs above 32,768 require the documented YaRN configuration and a compatible runtime.

Qwen2.5 license differences

Describing Qwen2.5 as “Apache 2.0 except for large models” is incorrect. The 3B checkpoint is the clearest counterexample because it uses the Qwen Research License despite being one of the smaller models.

• 0.5B, 1.5B, 7B, 14B, and 32B: Apache 2.0 in the main repositories.
• 3B: Qwen Research License, intended for non-commercial research and evaluation under its stated terms.
• 72B: custom Qwen License with additional provisions.

Always read the license file attached to the exact Base, Instruct, quantized, or derivative repository. Do not infer commercial permission from the family name.

Qwen2.5 context configuration

For 7B, 14B, 32B, and 72B, the official long-context instructions use YaRN with an original limit of 32,768 and a scaling factor of 4.

{
  "rope_scaling": {
    "factor": 4.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  }
}

Static YaRN applies the scaling factor even to shorter inputs and can affect short-context performance. Enable it only when the application needs inputs above 32,768 tokens.

Which Qwen2.5 models support about one million tokens?

One-million-token support belongs to two dedicated checkpoints. It must not be copied to the standard Qwen2.5 family.

These Qwen/... repository IDs are not interchangeable with provider service IDs. Copy the exact hosted identifier from the current provider catalog; see the Qwen API guide.

Core Qwen2.5 is not multimodal

The seven core Qwen2.5 model sizes accept text and generate text. Other capabilities are delivered by separate model families:

• Qwen2.5-Coder: specialized text models for programming tasks.
• Qwen2.5-Math: specialized text models for mathematical tasks.
• Qwen2.5-VL: a separate vision-language family for image, video, and text inputs with text output.
• Qwen2.5-Omni: a separate omni-modal family with text, image, audio, and video inputs and text or speech outputs.

Adding the term “multimodal” to the core Qwen2.5 checkpoints would misrepresent their supported inputs.

Base versus Instruct

Run Qwen2.5-7B-Instruct locally

The following example stays within the default 32,768-token configuration. It does not silently enable the 131,072-token YaRN extension.

pip install "transformers>=4.37.0" accelerate

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Qwen/Qwen2.5-7B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": "Summarize the advantages and risks of long-context inference."
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=384)

new_tokens = outputs[0][inputs.input_ids.shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

No fixed GPU claim can be applied to this snippet. Memory use changes with precision, quantization, model size, input length, generated length, batching, and framework overhead.

Frequently asked questions

Is Qwen2.5 a multimodal model?

The core Qwen2.5 checkpoints are text-only. Multimodal capabilities belong to separate Qwen2.5-VL and Qwen2.5-Omni families.

Are all Qwen2.5 models licensed under Apache 2.0?

No. The 3B repositories use the Qwen Research License, and the 72B repositories use the custom Qwen License.

Does every Qwen2.5 checkpoint support 1M context?

No. The documented open-weight 1M checkpoints are the dedicated 7B-Instruct-1M and 14B-Instruct-1M repositories.

Official sources

• Qwen2.5 release announcement
• Official Qwen2.5 collection
• Qwen2.5-3B-Instruct card and license
• Qwen2.5-7B-Instruct card and YaRN instructions
• Qwen2.5-72B-Instruct card and license
• Qwen2.5-7B-Instruct-1M card
• Qwen2.5-14B-Instruct-1M card

For programming-focused checkpoints, see the Qwen2.5-Coder guide. For the following general generation, see Qwen3.

This is an independent technical reference and is not an Alibaba Cloud or Qwen service.