Deepseek has unveiled its latest multimodal AI model, V4.1-Flash, featuring 552 billion parameters while significantly reducing KV cache memory usage to just a quarter of its predecessor’s requirements, according to The Decoder.
Despite only 16 billion parameters being active per token, the model narrowly outperforms rivals Opus 5 and GPT-5.6 Sol on the DeepSWE coding benchmark. The V4.1-Flash is released under the MIT license, aiming to enable more cost-effective AI agents.
For Japanese markets, where AI-driven trading and analytics are rapidly evolving, Deepseek’s efficient model could lower operational costs and accelerate adoption of advanced AI solutions in FX and equities sectors.