NostraEmpire/mirror-qwen2.5-32b-instruct

TEXT GENERATIONPricing:Input $2.72 / Output $4.8Concurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

NostraEmpire/mirror-qwen2.5-32b-instruct is a 32.5 billion parameter instruction-tuned causal language model from the Qwen2.5 series, developed by Qwen. This model significantly enhances capabilities in coding, mathematics, and instruction following, while also improving long text generation up to 8K tokens and structured data understanding, including JSON output. It supports a full context length of 131,072 tokens and is multilingual, covering over 29 languages.

Loading preview...

Qwen2.5-32B-Instruct Overview

This repository hosts the instruction-tuned 32.5 billion parameter Qwen2.5 model, part of the latest Qwen large language model series. Developed by Qwen, this iteration builds upon Qwen2 with substantial improvements across several key areas.

Key Capabilities & Enhancements

  • Enhanced Knowledge & Reasoning: Significantly improved performance in coding and mathematics due to specialized expert models.
  • Instruction Following: Demonstrates marked improvements in adhering to instructions and generating long texts (up to 8K tokens).
  • Structured Data Handling: Better at understanding structured data like tables and generating structured outputs, particularly JSON.
  • Robustness: More resilient to diverse system prompts, enhancing role-play and chatbot condition-setting.
  • Long-Context Support: Features a full context length of 131,072 tokens, with generation capabilities up to 8,192 tokens. It utilizes YaRN for handling extensive inputs, though static YaRN in vLLM may impact shorter text performance.
  • Multilingual: Offers support for over 29 languages, including major global languages like Chinese, English, French, Spanish, German, and Japanese.

Architecture & Training

The model is a causal language model based on transformers, incorporating RoPE, SwiGLU, RMSNorm, and Attention QKV bias. It has 64 layers and 40 attention heads (GQA) for Q and 8 for KV. The model underwent both pretraining and post-training stages.

Usage

Users can quickly get started with the model using the provided Hugging Face transformers code snippet, which includes apply_chat_template for easy interaction. For deployment with long contexts, vLLM is recommended, with specific rope_scaling configurations for YaRN.

Finetunes
1 models