internlm/internlm3-8b-instruct

InternLM3-8B-Instruct Overview

InternLM3-8B-Instruct is an 8-billion parameter instruction-tuned model developed by InternLM, designed for general-purpose applications and advanced reasoning. A key differentiator is its ability to achieve strong performance on reasoning and knowledge-intensive tasks while being trained on only 4 trillion high-quality tokens, representing a significant reduction in training cost compared to similarly scaled models like Llama3.1-8B and Qwen2.5-7B.

Key Capabilities

• Enhanced Reasoning: Excels in complex reasoning tasks, outperforming several competitors on benchmarks like GPQA-Diamond (37.4) and MATH-500 (83.0).
• Deep Thinking Mode: Features a specialized mode for solving complicated reasoning tasks via long chain-of-thought processes, alongside a normal response mode for fluent user interactions.
• Cost-Efficient Performance: Achieves competitive or superior results with over 75% less training data than other LLMs of similar scale.
• Strong General Performance: Demonstrates high scores across various benchmarks, including CMMLU (83.1), DROP (83.1), HellaSwag (91.2), and AlpacaEval 2.0 (51.1 LC WinRate).
• Coding Proficiency: Shows solid performance in coding benchmarks like HumanEval (82.3 Pass@1) and LiveCodeBench (17.8).

Good For

• Applications requiring robust reasoning and problem-solving capabilities.
• Scenarios where efficient performance with reduced training costs is critical.
• Tasks benefiting from both quick, fluent responses and detailed, step-by-step analytical thinking.
• General conversational AI and instruction-following tasks.