DeepSeek's V4.1 Flash Reads a Million Tokens for Pennies

DeepSeek released V4.1 Flash on September 10, putting a 552-billion-parameter model on its API and on Hugging Face under an MIT license. The headline is not raw size. It is how little the model costs to run.

Despite the parameter count, V4.1 Flash activates only a fraction of its weights on any given request. It is a mixture-of-experts design that fires roughly 8 billion parameters when reading a prompt and about 16 billion when writing a response, according to DeepSeek's own documentation. The rest of the network sits idle until it is needed, which is what keeps inference fast and inexpensive.

What it costs

Priced at $0.15 per million input tokens and $0.60 per million output tokens during off-peak hours, the model undercuts most Western frontier offerings by a wide margin. Those rates double at peak times, and a cache hit costs $0.003 per million tokens, which rewards workloads that reuse the same long context repeatedly.

The context window runs to a million tokens, with output of up to 384,000 tokens in a single generation. DeepSeek also built in native image understanding, so the model reads pictures without a bolt-on vision module. Cryptobriefing and the technical write-up at CellCog both put the on-disk checkpoint at about 510 gigabytes, trained from scratch on 45 trillion tokens.

An unusual shape

Most large language models use a single decoder stack for both reading and writing. V4.1 Flash splits the job. It runs an asymmetric encoder-decoder, with separate pathways tuned for taking in a prompt and for producing text. That asymmetry is why the active-parameter count differs between input and output, and it is a notable departure from the architecture DeepSeek used in earlier releases.

The launch lands in a crowded week for open weights. It follows DeepSeek's reported move to buy 160,000 Huawei Ascend chips for a new data centre, and arrives alongside a run of smaller open models built to run on modest hardware, including OpenBMB's on-device MiniCPM5. The through-line is cost: open-weight labs are competing less on peak benchmark scores and more on how cheaply a capable model can be served at scale.

For developers, the practical appeal is a model they can download, run on their own hardware, and point at documents that stretch to hundreds of pages without paying frontier prices. Whether V4.1 Flash matches the best closed models on hard reasoning is a separate question, and one the independent benchmarks will settle over the coming weeks.