Sakhanda Wire
NVDA $153.73 -1.54% MSFT $506.30 -2.26% GOOGL $280.09 -1.28% META $619.28 -1.14% AMZN $256.62 -0.44%
← Back to the news

Deepseek V4.1-Flash reduces memory demands for AI agents

Deepseek has released V4.1-Flash, a multimodal model designed to lower the cost of processing long contexts. The model has 552 billion parameters and supports contexts of up to one million tokens. Deepseek says its KV cache requires about one quarter of the fast-memory space used by Deepseek-V4-Flash, while the portion stored on SSD or host memory is reduced to roughly one eighth. Compared with Deepseek-V1, the global cache size per token is 437 times smaller.

The model separates its language backbone into an encoder and decoder. It activates 8 billion parameters per token when reading input and 16 billion when generating text, which Deepseek says nearly halves input-processing compute. The main cache is also stored in FP4 rather than FP8.

V4.1-Flash was trained on 45 trillion text and image tokens. On DeepSWE v1.1, Deepseek reports a 74.2 percent score, narrowly ahead of Anthropic’s Opus 5 and OpenAI’s GPT-5.6 Sol, but it trails significantly on ProgramBench. The report notes remaining weaknesses in demanding scientific tasks and complex image analysis. Model files are available on Hugging Face under the MIT license, and the model is offered through an API at V4-Flash’s prices.

Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.

Originally published by The Decoder on

Read the original on The Decoder ↗

The reporting and the image belong to The Decoder and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news