Lumina-Ai-Official/Nexa-AI-4B-Instruct

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Nexa-AI-4B-Instruct is a 4 billion parameter instruction-tuned large language model developed collaboratively by Neura Tech AI and Lumina AI, built upon Qwen/Qwen3-4B-Instruct-2507. This model features a 262,144 token context length and excels in multilingual understanding, coding, mathematics, and logical reasoning. It demonstrates strong performance across various benchmarks, particularly in reasoning and agentic tasks, making it suitable for complex instruction following and tool calling applications.

Loading preview...

Nexa-AI-4B-Instruct: A Collaborative Multilingual LLM

Nexa-AI-4B-Instruct is a 4 billion parameter instruction-tuned large language model, a joint effort by Neura Tech AI and Lumina AI. It is based on the robust Qwen/Qwen3-4B-Instruct-2507 architecture, inheriting its impressive 262,144 token context length.

Key Capabilities & Features

  • Multilingual Proficiency: Strong understanding and generation in multiple languages, including English, Hindi, and Chinese.
  • Advanced Reasoning: Demonstrates high performance in mathematical and logical reasoning tasks, as evidenced by strong scores on AIME25, HMMT25, and ZebraLogic benchmarks.
  • Coding & Agentic Tasks: Excels in coding assistance and is optimized for tool calling and AI agent applications, outperforming its base model and other comparably sized models in LiveCodeBench and various TAU agent benchmarks.
  • Instruction Following: Designed for high-quality instruction following and general conversation.
  • Fine-tuned Alignment: Features fine-tuned alignment for helpful and coherent responses.

Performance Highlights

Nexa-AI-4B-Instruct shows significant improvements over its base model and other models in its class across several critical areas:

  • Reasoning: Achieved 47.4 on AIME25 and 80.2 on ZebraLogic, showcasing superior problem-solving abilities.
  • Coding: Scored 35.1 on LiveCodeBench v6 and 76.8 on MultiPL-E.
  • Agentic Tasks: Demonstrated strong results in BFCL-v3 (61.9) and TAU1-Retail (48.7).
  • Alignment: Posted high scores in Creative Writing v3 (83.5) and WritingBench (83.4).

Ideal Use Cases

This model is particularly well-suited for developers and researchers requiring a capable, multilingual LLM for:

  • Building AI assistants that need to handle complex instructions and engage in general conversation.
  • Applications requiring strong mathematical and logical problem-solving.
  • Developing AI agents that interact with tools or perform multi-step tasks.
  • Code generation and assistance in various programming contexts.
  • Multilingual applications demanding high-quality understanding and generation.

Finetunes
1 models