DeepSeek's new model sets a template for powerful LLMs that run lean
DeepSeek V4.1 Flash proves that just because you build a bigger model doesn't mean you need more GPUs to serve it
Chinese AI startup DeepSeek has unveiled an updated version of its Flash model, called Flash 4.1, which promises to be a more powerful and resource-efficient large language model (LLM). The new model is 2.5 times larger than its predecessor, boasting 763 billion parameters. Despite its size, Flash 4.1 requires significantly less memory than larger models, thanks to several architectural improvements.
Two key advancements contribute to the model's efficiency. Firstly, DeepSeek modified how it handles key-value (KV) caches, which track model state across multiple sessions. By optimizing these caches, the model's memory requirements are reduced by up to 25 percent. Secondly, the introduction of N-gram parameters has enabled the model to process prompts more effectively while cutting down on computational and memory resources.
N-grams are groups of tokens, such as trigrams (three tokens) or bigrams (two tokens), that serve as a source of implicit knowledge or ingrained memory. This allows the model to quickly surface relevant information through cheap lookups, enhancing its intelligence without increasing performance penalties.
The N-gram parameters found in Flash 4.1 are similar to a technology initially developed by Google's Gemma team called Per-Layer Embedding (PLE). These parameters act as a "conditional memory module," decoupling memory from computation and enabling the model to be smarter while reducing the required compute and memory resources. This innovative approach could pave the way for larger, smarter, and less resource-intensive models in the future.
Written by urgent.news from The Register Science's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
