DeepSeek发布V3.2和V3.2-Speciale模型，对标GPT-5和Gemini3Pro

Two Models, Double the Impact

　　The company introduced two variants:

• V3.2 Standard Edition: Performs neck-and-neck with OpenAI's GPT-5 when handling documents up to 128,000 words
• V3.2-Speciale: Matches Google's Gemini3Pro on academic benchmarks while producing more detailed answers

Technical Breakthroughs Under the Hood

　　The secret sauce? A clever innovation called Directory-Style Attention (DSA). Traditional AI models struggle with long documents because processing time grows exponentially with length. DSA changes this dramatically:

• Makes processing time grow linearly instead of exponentially
• Uses 40% less memory
• Runs inferences 2.2 times faster

　　The result? These are the first open-source models capable of handling million-token documents on a single graphics card.

Smarter Thinking Through Better Training

　　The DeepSeek team didn't cut corners on training either:

• Dedicated over 10% of their computing power specifically to reinforcement learning
• Used group-based reinforcement learning (GRPO) combined with majority voting
• Removed artificial limits that discouraged lengthy reasoning chains

　　The payoff shows in testing - Speciale produces answers that are not only longer (32% more tokens than Gemini3Pro) but also more accurate (4.8 percentage points higher).

Open Source Commitment Continues

　　Both models are available now on GitHub and Hugging Face under the business-friendly Apache 2.0 license. DeepSeek promises more openness ahead:

"We're planning to release our DSA kernel and RL training framework next," a company spokesperson said.

　　The move continues DeepSeek's strategy of turning proprietary advantages into community assets - an approach that could reshape the competitive landscape by 2026 if they maintain this pace.

Key Points:

1. Performance Parity: Matches GPT-5/Gemini3Pro capabilities in respective domains
2. Technical Innovation: DSA enables efficient million-token processing
3. Training Investment: Significant computing resources devoted to RL optimization
4. Open Philosophy: Full weights available commercially under Apache 2.0 license