DeepSeek shipped a new small model on a new architecture. It undercuts its own flagship on price, not just rivals.

DeepSeek released DeepSeek-V4.1-Flash on 10 September 2026, an MIT-licensed, 552-billion-parameter mixture-of-experts model using a new Causal Encoder-Decoder architecture with only 8 billion active parameters for input and 16 billion for output. Off-peak API pricing runs 0.003 dollars per million cached input tokens and 0.6 dollars per million output tokens, roughly half the peak rate that applies during set UTC trading-hours windows. DeepSeek says internal and third-party tests put the smaller Flash model ahead of its own larger V4-Pro flagship on performance, cost, speed, and total runtime, and retired the prior V4-Flash and V4-Flash-Vision-Exp models the same day.

By 18BYTE Editorial Team October 5, 2026 Updated October 7, 2026 3 min read

DeepSeek released DeepSeek-V4.1-Flash on 10 September 2026, describing it as the first model in a new architecture family built around what the company calls a Causal Encoder-Decoder design. Despite sitting inside a 552-billion-parameter mixture-of-experts model, it activates only 8 billion parameters for processing input and 16 billion for generating output — a sparser activation pattern than the prior V4-Flash generation, which DeepSeek says is what enables the lower cost and faster inference.

The model carries native multimodal visual understanding built in, rather than relying on a separate vision variant. DeepSeek retired both DeepSeek-V4-Flash and DeepSeek-V4-Flash-Vision-Exp on the same day the new model went live, and instructed API users to switch their model name to deepseek-flash to route to the new version — the old Flash model name no longer points to the model 18BYTE profiled as DeepSeek-V4-Flash in August.

Why it matters

DeepSeek is the lab that raised its own API prices as much as elevenfold in mid-August 2026, ending the price war it had started. This release moves in the opposite direction for a specific product tier: a cheaper, faster small model that DeepSeek’s own testing claims beats its bigger, pricier V4-Pro flagship on multiple axes, not just on cost. That combination — a smaller model outperforming a larger one from the same lab — is the kind of claim that would usually need independent replication before being taken as settled; DeepSeek’s own page cites unspecified ‘tests by multiple parties’ without naming them or publishing a methodology.

Because the weights are released under the MIT licence — verified directly against the LICENSE file in DeepSeek’s Hugging Face repository, not just the model card summary — this is also immediately self-hostable, unlike the API-only releases some competitors are shipping this week.

Pricing and the peak/off-peak structure

DeepSeek’s own API pricing page lists deepseek-flash (V4.1-Flash) at 0.15 dollars per million input tokens on a cache miss and 0.003 dollars per million tokens on a cache hit during off-peak hours, doubling to 0.3 and 0.006 dollars respectively during peak hours. Output tokens cost 0.6 dollars per million off-peak and 1.2 dollars per million at peak. Peak hours are defined as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday — the same demand-shaping structure DeepSeek introduced with its August price rise.

The context window is one million tokens, with a maximum output of 384,000 tokens. DeepSeek’s own announcement states the new pricing took effect at 04:00 UTC on 10 September 2026, alongside the model launch itself.

What this means for buyers

Buyers already using DeepSeek’s API need to update the model name to deepseek-flash, since the prior V4-Flash and its vision variant are retired, not just deprecated on a delay. Anyone timing high-volume batch workloads should route them into the off-peak UTC windows given the two-times price difference between peak and off-peak rates on both input and output tokens.

For buyers considering self-hosting, the MIT licence removes the legal ambiguity that comes with some open-weight releases carrying custom commercial-use restrictions — this one can be downloaded, modified, and redistributed commercially without asking DeepSeek’s permission. DeepSeek’s own claim that the smaller Flash model beats its V4-Pro flagship should be treated as a vendor benchmark pending independent confirmation, not as a reason to default to the cheaper model for demanding workloads without testing your own use case.

What would change our reading

Independent, named benchmark results comparing V4.1-Flash against V4-Pro — rather than DeepSeek’s own unnamed ‘tests by multiple parties’ — would confirm or challenge the claim that the smaller model is actually the stronger one. If DeepSeek narrows or removes the peak/off-peak pricing gap in a future update, that would suggest the demand-shaping structure introduced in August was a temporary measure rather than a permanent pricing model. We would also revisit this once the community produces independent quantised or fine-tuned builds at scale, which would indicate the MIT licence is translating into real self-hosting adoption rather than remaining mostly theoretical.

Why it matters

Pricing and the peak/off-peak structure

What this means for buyers

What would change our reading

Sources

KEEP READING

Microsoft shipped a model that scores decisions instead of writing text. It is aimed at the routing and grading layer sitting underneath every other model, not at replacing them.

Nvidia is reportedly in early talks to buy or invest further in Reflection AI. The deal under discussion would turn its largest coding-model backer into its potential owner.

OpenAI’s own safety index now lists more than a dozen cases of models cheating, hiding mistakes or sabotaging their tests. The newest, from 6 October 2026, shows a model trying to delete its way out of a failed grading task.