DeepSeek v4.1 Flash

DeepSeek unveiled its V4.1 Flash model, asserting it outperforms its predecessor and delivers enhanced efficiency. The Chinese AI firm explained that the model leverages a novel "Causal-Encoder-Decoder" architecture built on a massive 552 billion-parameter framework. It adopts a Mixture-of-Experts (MoE) design, routing tasks solely to the subnetworks best suited for the job.

By activating just 8 billion parameters for processing inputs and 16 billion to generate responses, DeepSeek claimed to cut down computing power per request significantly. V4.1 Flash boasted native multimodal visual understanding, with a smaller footprint as the smallest model in DeepSeek's new series. Benchmarks revealed that V4.1 Flash surpassed V4 Pro in coding, cybersecurity, and autonomous agent tasks.

On Terminal-Bench 2.1, which evaluates AI on real-world compute tasks, V4.1 Flash scored 90.6, outpacing OpenAI's GPT-5.6 Sol (88.8), Moonshot AI's Kimi K3 (88.3), and DeepSeek's V4 Pro (87.9). In cybersecurity tasks on Cybergym, V4.1 Flash achieved 88.1, leading V4 Pro (83.3), Kimi K3 (80), and OpenAI's GPT-5.6 Sol (84.5). V4.1 Flash narrowly edged out Anthropic's Claude Opus 5 on a software engineering benchmark (0.02 yuan per million cached input tokens during off-peak hours).

DeepSeek's head of the Harness team, Cui Tianyi, stated that the model comprehensively outperformed V4 Pro in all metrics, and the company would route V4 Pro requests to the more affordable Flash model temporarily until a more powerful V4.1 Pro was ready.