DeepSeek V4.1 Flash: Performance and Architecture Overview

Architecture and Efficiency

DeepSeek V4.1 Flash utilizes a 552-billion-parameter mixture-of-experts backbone supplemented by 196 billion parameters in Engram memory modules. The model is designed for efficiency, activating approximately 8 billion parameters during input processing and 16 billion during generation. A new causal encoder-decoder design allows input tokens to bypass full decoder computation, which reduces GPU memory usage for the global attention cache by roughly 75% compared to previous generations.

Performance and Benchmarking

When configured with maximum thinking effort, the model achieved an 81.25% score on the KingBench 3 coding suite, a significant improvement over the 53.75% score recorded when thinking was disabled. The model demonstrated high proficiency in complex tasks, including a restricted-moves math problem (solved correctly as 20,460) and a local Gemma 2B fine-tuning workflow that produced a functional web application. Generation speeds averaged approximately 221 tokens per second, inclusive of reasoning tokens. While the model shows strong capabilities in logic and simulation, it still exhibits minor edge-case bugs in UI-heavy tasks like the folding table simulation and archery game result screens.

API and Usage

DeepSeek V4.1 Flash is priced at $0.15 per million uncached input tokens and $0.60 per million output tokens during off-peak hours. Peak pricing is $0.30 and $1.20 respectively. Cached input tokens are priced at $0.003 per million during off-peak hours. Users are advised to explicitly enable thinking and set effort to maximum within their API configuration to achieve the performance levels observed in these tests.