DeepSeek V4.1-Flash Ships 552B Open Weights Under MIT

What DeepSeek Released on September 10

DeepSeek published DeepSeek-V4.1-Flash on September 10, 2026, and put the weights on Hugging Face under the MIT licence. It is the smallest model in the lab's new architecture family, and the interesting part is not that it is small — it is that DeepSeek says this small model now beats the larger V4-Pro it is replacing. The company is retiring V4-Pro into it four days after launch.

• 552 billion total parameters, with only 8 billion active per token while reading a prompt and 16 billion while writing
• 1 million tokens of context and native image-and-text input, per the model card
• MIT licence on the weights, so anyone can download, modify and redistribute the model
• $0.15 per million input tokens and $0.60 per million output tokens at off-peak rates, with peak rates exactly double

DeepSeek's changelog states that from September 14 at 04:00 UTC, every request sent to the older deepseek-v4-pro model is routed to V4.1-Flash and billed at the Flash tier. A vendor deprecating its own flagship into a cheaper model is an unusually direct claim, and it is worth reading it as exactly that: DeepSeek's own assessment of its own product.

How the Causal Encoder-Decoder Architecture Saves Memory

The architecture is where this open-weight model earns its name. V4.1-Flash uses what DeepSeek calls a Causal Encoder-Decoder layout: 40 transformer layers split into a 20-layer causal encoder followed by a 20-layer decoder. Instead of every layer maintaining its own key-value cache, the decoder builds its cache once from the encoder's output.

The practical result is memory. DeepSeek reports a KV cache cost of roughly 890 bytes per token — about a quarter of what the previous V4-Flash required at the same context length. On a million-token context, that difference is the gap between a workload that fits on the hardware you have and one that does not. If you are sizing a machine for this kind of model, our guide to how much RAM a local LLM needs walks through the arithmetic.

The mixture-of-experts routing is aggressive too: 384 routed experts per MoE layer with six activated per token. That asymmetry — 552 billion parameters resident, 8 billion doing work on the read path — is what keeps inference cheap enough to justify the pricing.

What Does DeepSeek V4.1-Flash Cost to Run?

DeepSeek's published API pricing puts cache-miss input at $0.15 per million tokens and output at $0.60 per million tokens during off-peak hours, with cached input at $0.003 per million. Peak hours, which DeepSeek defines as 01:00–04:00 and 06:00–10:00 UTC on weekdays, cost twice that. Everything outside those windows is off-peak.

Because the weights are MIT-licensed, the API price is a ceiling rather than the only option. Teams with their own hardware can serve the model themselves, and the reduced KV cache footprint is what makes that a realistic conversation for a million-token context rather than a theoretical one.

Why Open Weights at This Size Matter

DeepSeek publishes benchmark figures alongside the release: 90.9 on GPQA Diamond, 90.6 pass@1 on Terminal-Bench 2.1, 74.2 resolved on DeepSWE v1.1, and a Codeforces rating of 3471. Those are the lab's own numbers, measured with its own harness, and TNW notes that none of them have been independently verified — so treat them as claims to be checked rather than settled results. DeepSeek's own comparison tables also show the model trailing frontier reasoning models on the hardest evaluation sets.

What is not in dispute is the licence and the file. A 552-billion-parameter multimodal model with a million-token window, released under MIT, is downloadable today by any researcher, startup or hobbyist who wants it. That pattern has been the story of open weights all year — K2 Horizon shipped six open models with their training data less than a week ago — and each release moves the floor of what a self-hosted system can do. More on where open models are heading in our AI coverage.

Sources: DeepSeek API Docs changelog — September 10, 2026; DeepSeek-V4.1-Flash model card on Hugging Face — September 2026; DeepSeek API pricing — September 2026; TNW — September 10, 2026.