# M3 一次可處理高達 100 萬個標記，為前代的 5 倍。

*genai · news · 2026-06-01 · Sputnik News*

## Key points

- M3 一次可處理高達 100 萬個標記，為前代的 5 倍。
- M3 在 SWE-Bench Pro 測試中取得 59% 分數，超越 GPT-5.5 與 Gemini 3.1 Pro。
- M3 的 Sparse Attention 架構將運算需求降低最多 95%，成本減少 90%。
- M3 在基準測試中自主將 NVIDIA Hopper 晶片利用率從 7.6% 提升至 71.3%。

M3 processes up to 1 million tokens at once - 5x more than its predecessor, enabling it to handle massive codebases The model scored 59% on SWE-Bench Pro, outperforming OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro in real-world software engineering tests Its new Sparse Attention architecture cuts computing requirements to as little as 1/20th of previous levels, reducing costs by over 90% while enhancing speed In one benchmark, M3 autonomously optimized software for NVIDIA Hopper chips, boosting hardware utilization from 7.6% to 71.3%

**Companies:** NVIDIA

[Read the full story on Sputnik News](https://sputnikglobe.com/20260601/how-chinas-new-ai-model-beating-openai--google---1124235849.html)

---

Canonical: https://newsio.io/zh-TW/n/33a0b9ba-c5f4-4072-b56e-c5eed9fc2555/m3-100-5-sparse-attention-120
Summarized by Newsio from Sputnik News. https://newsio.io/how-it-works