DeepSeek V4 Lands Open, MIT-Licensed, With a Million-Token Context
DeepSeek released the V4 public preview on Friday - a 1.6-trillion-parameter mixture-of-experts flagship with just 49 billion active per token, a 284B Flash sibling, native million-token context, and MIT licensing on Hugging Face. The open frontier now ships capabilities the closed labs charge premiums for, days after GPT-5.5 made the million-token window its headline feature.
Three days after OpenAI shipped GPT-5.5 with a million-token context window as its marquee feature, DeepSeek released the same capability under an MIT license, downloadable from Hugging Face. The V4 public preview, which landed Friday, is a two-model family: V4-Pro, a 1.6-trillion-parameter mixture-of-experts flagship that activates just 49 billion parameters per token, and V4-Flash, a 284-billion-parameter sibling activating 13 billion — both with native million-token context, output lengths up to 384K tokens, and weights shipped in FP8 and FP4 mixed precision for cheap serving.
The timing reads as it always reads with DeepSeek: the closed frontier announces a capability as a premium; the open frontier ships it as a commodity inside a news cycle. What is worth analyzing is not that this happened — it is the fourth or fifth iteration of the pattern — but what is different about this iteration: the sparsity ratios, the licensing choice, and what "preview" means when the artifact is already on Hugging Face.
The sparsity story is the cost story
The headline parameter count — 1.6 trillion — is marketing physics; the number that matters is 49 billion active. V4-Pro's roughly three-percent activation ratio means the model carries frontier-scale capacity while paying inference costs closer to a mid-size dense model, and V4-Flash pushes the same logic further: 13 billion active parameters is small enough to serve economically at high volume on previous-generation hardware. Add the FP4-ready weights and the release is engineered, top to bottom, around a single thesis: that the winner of the open-model race is decided by cost per served token, not by leaderboard position. Community-reported serving prices are already circulating well below a dollar per million output tokens; those figures are early and thinly sourced, but the architecture makes their direction credible.
This is the same wager visible across the Chinese open-weight ecosystem this month — Zhipu's GLM-5.1 shipped MIT-licensed at frontier-adjacent capability last week — and it is the pincer movement I described in last week's analysis of the open-source frontier squeeze: closed labs are being compressed between capability parity above and cost collapse below, and V4 advances both jaws at once.
MIT, and why the license is the strategy
V4 ships under straight MIT — no acceptable-use policy, no field-of-use restrictions, no revocation mechanics. For enterprise legal departments, that distinction is not ideological; it is the difference between a dependency and a relationship. An MIT artifact can be self-hosted, fine-tuned, embedded, and resold without ongoing permission from a Chinese vendor — which neutralizes, by construction, most of the procurement objections that keep Chinese-origin models out of Western enterprises. The license is doing geopolitical work: it converts "can we depend on DeepSeek?" into "can we depend on this file?", and files do not have export policies. Whether Washington eventually decides that files, too, are a policy surface is one of the more consequential open questions of the year.
The million-token context deserves the same reading. Long context is the enabling capability for the workloads enterprises actually want — whole- codebase reasoning, full-archive retrieval, the document-heavy agent pipelines that are becoming the default enterprise ambition. Making it native and open means the capability ceiling for self-hosted deployments just rose to match what the closed APIs sell, days after the data-wall debate suggested capability itself is getting harder to differentiate.
What "preview" means here
The label matters less than the artifact. A closed lab's preview is a waitlist; an open lab's preview is version one in your infrastructure. The weights are downloadable now, which means fine-tunes, quantizations, and serving optimizations begin this weekend, and the stable release — whenever DeepSeek stamps it — will arrive into an ecosystem that has already absorbed the model. That absorption is the real moat the open ecosystem is building: not any single model, but the compounding infrastructure of tooling, hosts, and organizational familiarity that makes the next open release cheaper to adopt than the last.
What to watch from here: whether US inference providers list V4 at the aggressive prices the architecture enables, and how quickly; whether the frontier labs respond on price for long-context workloads specifically, where their premium is now most exposed; and what the first independent benchmark sweeps show once the community stops benchmarking the preview and starts benchmarking its fine-tunes. The pattern of the last two years says the fine-tunes are where open releases actually compete — and with MIT licensing and three-percent activation, V4 was built to make that competition as cheap as possible.