Qwen2.5 Math 1.5B Instruct
A 1.5B-parameter Qwen2.5-Math chat model that solves English and Chinese math problems step by step and boxes the final answer. Qwen reports 79.7 on MATH for it when an interpreter runs the Python it writes.
Overview
Qwen2.5 Math 1.5B Instruct is the smallest chat model in Qwen's Qwen2.5-Math series. It is fine-tuned from the Qwen2.5-Math-1.5B base model on solutions picked by Qwen's 72B math reward model, followed by GRPO reinforcement learning with that reward model.
It answers English and Chinese math problems in two modes. Chain-of-thought (CoT) writes the derivation in words and formulas. Tool-integrated reasoning (TIR) mixes the explanation with Python programs that handle exact arithmetic. Qwen reports 79.7 on MATH for this model with TIR and says it surpasses most earlier 70B math models.
Choosing a variant: Qwen2.5 Math 7B Instruct is the larger sibling, at 83.6 on MATH with CoT and 85.3 with TIR in Qwen's results. Start with this 1.5B model for exercises at the level of the default prompt, a one-variable linear equation, and move up to 7B if it gets problems from your own set wrong.
What you can build
- A homework helper for school algebra that shows every step and ends in one boxed answer your app can extract.
- Self-consistency voting: run one problem with several seeds and keep the most frequent answer.
- Chinese or English practice drills that reveal the worked solution once the learner has answered.
Inputs
The only required field is the prompt: the problem as text. Inline LaTeX works, as in the default prompt, which asks for the value of x that satisfies 4x+5 = 6x+7. An optional system prompt goes before it. Fill in user_id and session_id to keep chat history, so a learner can ask about one step later.
Everything else tunes decoding: temperature (0 to 5, default 0.7), top_p (default 0.95), top_k, repetition and length penalties, token limits, semicolon-separated stop sequences and a seed (default 123456). Two switches, do_sample and quantization, are both on by default.
Outputs
Plain text with LaTeX. For the default prompt the sample works in four numbered steps: restate the equation, subtract 4x, subtract 7, divide by 2. It closes with -1 inside \boxed{}, which renders as a framed answer.
Recommended settings
- Replace the default system prompt with Qwen's CoT instruction: "Please reason step by step, and put your final answer within \boxed{}." Qwen's TIR prompt asks for natural language reasoning combined with programs.
- For one repeatable answer, set do_sample to off. For voting, Qwen's own evaluation samples at temperature 0.7 and top_p 0.8.
- If you cap max_tokens, leave room for the whole derivation so the boxed answer is not cut off.
Limitations
- Qwen does not recommend the series beyond English and Chinese math, so route general chat and coding elsewhere.
- In TIR mode the model writes Python that Wiro returns as text without running. Qwen's TIR scores assume the code runs, so you need your own sandbox.
- Qwen notes that CoT alone has trouble with exact computation, giving quadratic roots and matrix eigenvalues as examples.
- Context is short (Qwen pre-trained the series at 4K), so keep session history trim.
- It scores below the 7B and 72B models on MATH, so check the boxed answer before a student sees it.
Safety & compliance
Qwen releases this model under the Apache 2.0 license and states its intended use as solving English and Chinese math problems.
API quick start
Run Qwen2.5-Math-1.5B-Instruct with a single API call.
{
"prompt": "Find the value of $x$ that satisfies the …",
"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-math-1-5b-instruct" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Find the value of $x$ that satisfies the …",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



