DeepSeek Releases V4.1 Flash With Million-Token Context
DeepSeek has launched V4.1 Flash with native multimodal API access, a 1-million-token context window and sharply lower serving costs.
DeepSeek has released V4.1 Flash through its API, introducing a new architecture aimed at reducing the memory and serving burden of long-context inference. The company says the model combines a 552-billion-parameter mixture-of-experts backbone with a causal encoder–decoder design, activating roughly 8 billion parameters during input processing and 16 billion during decoding.
A cheaper long-context proposition
The release keeps DeepSeek’s focus on making large-model capability affordable in production. DeepSeek says V4.1 Flash reduces key-value cache requirements to one-quarter of the previous generation’s HBM usage and one-eighth of its SSD storage requirement. The model supports a 1-million-token context window and is available through the API with native multimodal support.
DeepSeek is also passing part of the efficiency gain to customers. Its published launch pricing sets output at 4 yuan per million tokens during off-peak periods, with peak pricing at twice that level. Cached input is priced at 0.02 yuan per million tokens, while uncached input costs 1 yuan. The company has also said it will work with the open-source community on inference support and broader deployment options.
A forced product reset
The launch changes DeepSeek’s product hierarchy as well as its price sheet. V4 Flash and V4 Flash Vision Experimental have been retired, while requests aimed at V4 Pro will be routed to V4.1 Flash after September 14 until a successor Pro model arrives. DeepSeek says internal and external testing showed the new Flash model exceeding V4 Pro on performance, cost, speed and total completion time, although those comparisons have not yet been independently audited.
Why it matters
This is significant because the release targets the most expensive part of practical AI deployment: sustained, very long-context workloads. If the cache reductions survive independent testing, developers may be able to run document, coding and agent systems with less accelerator memory and storage overhead. The unresolved question is whether the claimed efficiency translates cleanly beyond DeepSeek’s own serving stack and benchmark conditions.