Qwen2.5 3B Instruct
The 3.09B-parameter Qwen2.5 chat model that Qwen added for mobile applications. Its license, the Qwen Research License, allows non-commercial research and evaluation only.
Overview
Qwen added a 3B size with Qwen2.5 after finding user interest in 3B models for mobile applications; Qwen2 had nothing between 1.5B and 7B. The instruction-tuned checkpoint has 3.09B parameters (2.77B non-embedding), 36 layers, tied embeddings and a 32,768-token context, and Wiro accepts about that much input per request.
Qwen's table compares it with the smaller Gemma2-2B-IT and two slightly larger models, Phi3.5-mini-Instruct and MiniCPM3-4B. It leads both on MATH (65.9 against 48.5 and 46.6) and MultiPL-E (60.2 against 47.2 and 49.1), and trails MiniCPM3-4B on IFEval strict-prompt (58.2 against 68.4).
Choosing a variant: its license limits the 3B to non-commercial research and evaluation. For a commercial product, Qwen2.5 1.5B Instruct and Qwen2.5 7B Instruct are Apache 2.0, and the 7B outscores the 3B on every instruct test Qwen reports for both. Qwen2.5 0.5B Instruct covers label-style tasks and Qwen2.5 32B Instruct the heaviest math and code prompts. Qwen2 7B Instruct predates all of them.
What you can build
- A feasibility test with your real prompts before committing to a 3B-class on-device assistant.
- Math tutoring experiments with step-by-step solutions (Qwen reports 86.7 on GSM8K).
- Code generation tests in several programming languages, the skill MultiPL-E measures.
- Research on small-model JSON extraction or function calling.
Inputs
A prompt or a messages array carries the request; send one of the two. The default system prompt, a generic instruction to be helpful and avoid harmful content, can be overwritten.
Pass a user ID and a session ID, and Wiro replays that session's history. About 32,768 input tokens cover the whole request, tools and history included. Beyond that, the oldest history is dropped first and then the tail of the newest message. Function definitions are kept whole, and a request whose definitions and system prompt leave no room for the message fails with a context-length error.
Also available: up to 128 function definitions, a tool-choice setting, parallel tool calls (on by default), JSON mode, and the usual sampling, length, stop and seed controls.
Outputs
Output is text unless JSON mode or a tool call applies. Leaving max tokens at 0 gives each reply a budget of 2,048 tokens, or twice the prompt length when the prompt itself runs past 2,048 tokens. A value you set is honoured within the room left in the window. The sample gathers Eiffel Tower facts, including its planned dismantling after the 1889 Exposition and a 330-meter height. Its figures, such as more than 70 lightning strikes, need checking before reuse.
Recommended settings
- Set top-k to 20, top-p to 0.8 and the repetition penalty to 1.05, as in the checkpoint's generation config. Temperature 0.7 already matches.
- Keep instructions short and explicit, and use JSON mode with sampling off for extraction tests.
Limitations
- Complex formatting instructions can be missed, going by the 58.2 IFEval score.
Safety & compliance
The Qwen Research License limits use to non-commercial research or evaluation, and commercial users must request a license from Alibaba Cloud. If you use the model or its outputs to build, train or improve an AI model that you distribute or make available, its documentation must display "Built with Qwen" or "Improved using Qwen".
API quick start
Run Qwen2.5-3B-Instruct with a single API call.
{
"prompt": "Can you say something interesting about t…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/qwen/qwen2-5-3b-instruct" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Can you say something interesting about t…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.
