Key takeaways

• DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts (MoE) model designed for fast inference and strong performance, leveraging 284B total and 13B activated parameters.
• The model integrates hybrid compressed attention and improved residual connections, supports a 1M-token context window, and is open-weights under the MIT license.
• It excels in mathematical problem-solving, achieving 97.4% accuracy on the MATH-500 dataset, particularly in arithmetic, algebra, and formula application. However, it struggles with strict output formatting and subtle problem interpretations in math.
• DeepSeek V4 Flash demonstrates significantly weaker performance (32.7%) in complex multi-turn text-to-SQL tasks on the BIRD-CRITIC dataset. Its limitations in text-to-SQL include difficulties with advanced window functions, temporal logic, intricate set-based operations, and dynamic schema interaction.
• Key error modes in text-to-SQL tasks involve incorrect window function usage, temporal logic errors, issues with complex JSON/JSONB manipulation, and failures in handling constraints and triggers.

Technical specifications

Context length

1M tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~284B

Open weights

Yes

Licence

MIT

Regions

China

Vendor

DeepSeek

Provider model key

deepseek/deepseek-v4-flash