DeepSeek just dropped a model that processes a million tokens of context while activating fewer parameters than some open-source models released two years ago. The V4.1-Flash, launched on September 10, represents the Chinese AI startup’s latest bid to rewrite the economics of large language models.
The model packs 552 billion parameters into a Mixture-of-Experts (MoE) architecture, but only fires up about 8 billion of them for input tasks and 16 billion for output. The result is a model that punches well above what its active compute footprint would suggest.
The architecture that makes it work
V4.1-Flash introduces what DeepSeek calls an asymmetric Causal Encoder-Decoder architecture, processing input and generating output through different pathways optimized for each task, rather than running everything through a single pipeline.
The context window stretches to 1 million tokens. Supporting that massive context is a KV cache that consumes approximately 890 bytes per token, about one-quarter of what the prior V4-Flash model required.
