What DeepSeek V4 is
DeepSeek V4 is the flagship AI model from DeepSeek, a Chinese AI research lab. It is a large language model — the kind of AI that can read, write, reason, and code — and it comes in two versions: V4-Pro, a very large model built for maximum capability, and V4-Flash, a smaller, faster version that trades a little performance for lower cost and quicker responses.
What sets DeepSeek V4 apart from most frontier AI models is that it is open-weights: DeepSeek publishes the full model files on Hugging Face, so anyone can download, inspect, and run it. That's unusual at this level of capability, where most comparable models (like GPT-5.5 or Claude Opus 4.8) are only available through paid APIs.
Why it matters — even if you've never heard of DeepSeek
If you use AI tools at work, DeepSeek V4 matters for a simple reason: it has pushed prices down across the whole industry. DeepSeek permanently cut V4-Pro API prices by 75%, continuing a pattern where each new DeepSeek release forces competitors to reconsider what they charge. The original DeepSeek-V3 was already priced at $0.27 per million input tokens — a fraction of what Western labs charged at the time.
For developers, the open-weights release means you can run V4 on your own infrastructure, customize it for your use case, and avoid depending on any single vendor's API.
How it works (the basics)
DeepSeek V4 uses an architecture called Mixture-of-Experts (MoE). Here's the plain-English version: instead of one giant brain that processes everything, the model is divided into many specialized sub-networks ("experts"). For any given input, only a small fraction of those experts are activated. V4-Pro has 1.6 trillion parameters in total, but only 49 billion are "switched on" for each token it processes. V4-Flash is smaller — 284 billion total, 13 billion active.
This design is what makes the model both powerful and practical to run. It also enables a key feature: both variants support a 1 million token context window by default. In practical terms, that means you can feed the model an entire large codebase, a long legal document, or hours of meeting transcripts in a single request. This is enabled by a new attention technique called DeepSeek Sparse Attention (DSA), which reduces the compute cost of handling very long inputs.
The road to V4: a fast-moving lineage
DeepSeek V4 didn't appear from nowhere. The lab has shipped a rapid series of models leading up to it:
- DeepSeek-V3 introduced the MoE approach at scale with 671 billion parameters, running at 60 tokens per second — three times faster than its predecessor — and was fully open-sourced.
- DeepSeek-R1 added reasoning capabilities on par with OpenAI's o1, also under a permissive MIT license.
- V3.1, V3.1-Terminus, V3.2, and V3.2-Exp iterated on agent capabilities, tool use, and the sparse attention architecture that V4 now ships with by default.
- V4 brings all of this together: 1M context, two model sizes, open weights, and API compatibility with both OpenAI and Anthropic formats.
The controversy: where DeepSeek's training data comes from
DeepSeek's rapid progress has attracted serious scrutiny. In February 2026, Anthropic publicly accused DeepSeek — along with Moonshot AI and MiniMax — of running coordinated, large-scale attacks against Claude. According to Anthropic, approximately 24,000 fraudulent accounts were used to generate over 16 million AI conversations, specifically targeting Claude's most distinctive capabilities like reasoning, tool use, and coding. Anthropic framed this as a national security concern, arguing that models trained this way inherit capabilities without the safety guardrails that went into the original.
A separate investigation detailed a broader gray-market ecosystem of API proxy services giving Chinese developers discounted access to U.S. AI models, with call logs harvested as training data. The White House acknowledged the distillation threat in an April 2026 memo.
DeepSeek has not publicly addressed these specific allegations in the events covered here.
Recent developments
- DeepSeek permanently cut V4-Pro API prices by 75% in May 2026 — a sustained pricing move, not a promotion.
- V4-Pro and V4-Flash were released on Hugging Face in April 2026, with V4-Pro accumulating over 4.3 million downloads quickly.
- Legacy DeepSeek model endpoints are scheduled to be retired in July 2026, with V4 as the replacement.
- Industry analysis notes that V4 trails leading open and closed models on some aggregate benchmarks, even as it claims open-source state-of-the-art on agentic coding tasks.
Where it's heading
DeepSeek's trajectory points toward continued cost pressure on the industry and deeper investment in agentic capabilities — models that can take multi-step actions, use tools, and run autonomously. The open-weights strategy has also inspired community reproductions and fine-tunes, extending the model's reach well beyond DeepSeek's own API. Whether the geopolitical and safety questions around its training data lead to regulatory constraints remains an open question.
