Chinese LLMs tested for speed, reveal output token count as key factor

A developer's attempt to highlight the slowness of Chinese LLMs revealed unexpected performance characteristics across various models. While many Chinese models like Kimi K3, Qwen 3.8 Max, and MiniMax M3 were found to be significantly slower than Western counterparts such as Claude 5 Opus, even when handling large contexts, the author discovered that output token count, rather than raw processing speed, was the primary driver of perceived slowness. Surprisingly, Claude 4.5 Haiku, expected to be fast, was also found to be among the slower models tested. AI

IMPACT Highlights that output token count significantly impacts perceived LLM speed, suggesting developers should consider this metric alongside raw processing power.

RANK_REASON The item is a personal blog post reflecting on LLM performance rather than an official release or benchmark.

Read on dev.to — LLM tag →

• Claude
• Claude 4.5 Haiku
• Claude 5 Opus
• DeepSeek Flash
• DeepSeek V4 Pro
• Gemini
• Gemini Pro
• GLM-5.2
• Grok
• Kimi K3
• MiniMax M3
• Mistral AI
• Mistral Large
• Mistral-medium
• OpenAI
• Qwen 3.7 Flash
• Qwen 3.7 Plus
• Qwen 3.8 Max
• Sakana Fugu base

• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →