Qwen2.5 1.5B Instruct

Overview

Qwen2.5-1.5B-Instruct is the post-trained chat version of the 1.54B-parameter Qwen2.5 base, with 1.31B non-embedding parameters in 28 layers and shared input and output embeddings. On Wiro, one request can carry about 32,768 input tokens, the same length as the card's context.

Qwen groups this size with the 0.5B as an edge-side model for tight resource budgets. Compared with Qwen2-1.5B-Instruct, the release post shows the biggest gains in math and code: MATH rises from 25.3 to 55.2 and HumanEval from 42.1 to 61.6. IFEval strict-prompt moves from 29.0 to 42.5.

Choosing a variant: this is the lightest model in the family that scores above 50 on both MATH and HumanEval. Drop to Qwen2.5 0.5B Instruct only for routing or tagging against fixed labels. Qwen2.5 3B Instruct beats it on all eleven instruct benchmarks Qwen lists for both, but its license is non-commercial. For open-ended chat, Qwen2.5 7B Instruct and Qwen2.5 32B Instruct fit better. Qwen2 7B Instruct belongs to the generation before.

What you can build

• Two-line summaries of support tickets or commit messages for an internal feed.
• Catalog tagging that reads a product blurb and returns its category and material as JSON fields.
• Worked answers to grade-school word problems, the type GSM8K measures (Qwen reports 73.2).
• First-draft replies in any of the more than 29 languages Qwen2.5 supports, for a person to review.

Inputs

The prompt field is required unless you send a messages array instead. The system prompt arrives filled with a stock instruction about honest, safe answers; swap in your own role or output format.

Wiro stores history when you pass a user ID with a session ID, and a new session ID starts fresh. System prompt, stored turns, tool definitions and your message share a budget of about 32,768 tokens. Over budget, Wiro removes the oldest turns, then shortens the end of the latest message. Function definitions are not shortened, so a request whose definitions and system prompt fill the budget fails with a context-length error.

Other fields cover JSON mode, function definitions, sampling controls (temperature 0.7 by default), minimum and maximum length, semicolon-separated stop strings and a seed of 123456 unless you change it. Quantization starts enabled.

Outputs

You get text back, JSON when you ask for it, or tool calls when functions are defined. If max tokens stays at 0, a reply may use 2,048 tokens for a prompt of 2,048 tokens or less and twice the prompt's length for a longer one, and a value you set applies as long as the window has room after the input. The page's sample for the default prompt is a four-sentence paragraph on the Eiffel Tower: a temporary iron structure for the 1889 World's Fair that became permanent, 324 meters tall.

Recommended settings

• Copy the sampling values from the checkpoint's generation config: temperature 0.7, top-p 0.8, top-k 20, repetition penalty 1.1.
• For tagging, request a JSON object, spell out the keys in the prompt and switch sampling off.

Limitations

• Qwen reports 14.8 on LiveCodeBench, so leave real programming work to the larger sizes.
• Broad knowledge is limited (50.7 on MMLU-redux). Check facts before showing them to users.
• The card lists generation up to 8,192 tokens, while a short prompt with max tokens left at 0 gets a 2,048-token reply budget, so raise the limit for long answers.

Safety & compliance

License: Apache 2.0. Commercial products and fine-tuned derivatives are both allowed.