New Deepseek model V4.1-Flash cuts memory needs for AI agents
decoder ·
DeepSeek has released V4.1-Flash, a multimodal model designed mainly to lower the operating costs of handling long contexts. The main improvement is a much smaller KV cache: the fast-GPU buffer needs about a quarter of the space used by DeepSeek-V4-Flash, the offloaded part shrinks to about an eighth, and the global KV cache per token is down by a factor of 437 compared with V1. The language model has 552 billion parameters and supports contexts up to one million tokens. It splits the backbone into an encoder and decoder, activating only 8 billion parameters per token when reading input but 16 billion when generating output, and it stores the main KV cache in FP4 instead of FP8, nearly halving that cache's memory footprint. DeepSeek trained it from scratch on 45 trillion text and image tokens and says post-training gains came from larger, better-controlled data, tasks, and environments rather than new algorithms. On coding benchmarks it can match top closed models, narrowly beating Anthropic's Opus 5 and OpenAI's GPT-5.6 Sol on DeepSWE v1.1 at 74.2 percent, but it still trails badly on ProgramBench and shows gaps on complex scientific tasks and image analysis. The model is available on Hugging Face under the MIT license and via API at the same prices as V4-Flash, with a user-adjustable thinking depth that improves results at the highest setting but produces about 2.5 times as many output tokens.
decoder