DeepSeek V3.2 Released: Breakthroughs in Long-Text Capability, API Prices Cut in Half

Overview

Recently, the domestic AI startup DeepSeek officially released its latest generation large language model, DeepSeek V3.2. This update not only achieved significant breakthroughs in core technical indicators such as long-text processing but also adopted an aggressively competitive pricing strategy in its commercialization approach, slashing API call prices by 50%. This move aims to lower the entry threshold for enterprise users, further consolidate its competitiveness in the field of domestic large models, and accelerate the implementation of Artificial General Intelligence (AGI) applications.

Technical Architecture and Long-Text Breakthroughs

DeepSeek V3.2 is built on a deeply optimized Mixture of Experts (MoE) architecture, significantly enhancing the model’s ability to understand and generate super-long contexts through the reconstruction of underlying logic. According to official disclosures, the V3.2 version supports a context window of up to 256K, and when processing documents spanning thousands of pages, complex codebases, or multi-turn long conversations, it maintains extremely high fidelity to information and logical consistency, effectively resolving the "lost in the middle" pain point present in previous versions.

• Performance Optimization: In mainstream evaluation benchmarks such as MMLU, HumanEval, and LongBench, V3.2’s comprehensive scores showed significant improvement over V3.0, particularly excelling in mathematical reasoning and complex instruction following.
• Multimodal Enhancement: The new model strengthens its ability to parse visual information, enabling more precise extraction of key data from charts and diagrams, providing stronger tool support for scientific research analysis and industrial quality inspection.

Pricing Strategy and Market Impact

In terms of market promotion, DeepSeek announced a "halving" of API prices. Specifically, the unit price for its online inference interface was reduced to half of the original price, while the price for cached inference interfaces (Cache API) dropped to as low as 20% of the original price. This radical pricing strategy is expected to have a profound impact on the domestic and global foundation model service market, forcing competitors to reexamine their own pricing systems.

"We hope to make high-quality large model capabilities accessible to all, like water and electricity, completely eliminating the cost barriers to technology use." said the DeepSeek R&D team.

Industry analysis indicates that this move not only reflects DeepSeek’s technical advantages in computing cost control but also demonstrates its strategic determination to accelerate the popularization of AI applications and cultivate the developer ecosystem. For the vast number of enterprises relying on large models for content generation, code assistance, and data analysis, this directly translates into a significant reduction in operating costs and an increase in business expansion speed.