What DeepSeek V4 is
DeepSeek V4 is the current flagship AI model family from DeepSeek, a Chinese AI research lab. It's a large language model — the kind of AI that can read, write, reason, and write code — and it comes in two sizes: V4-Pro (the more powerful version) and V4-Flash (faster and cheaper). Both are released as open-weights models, meaning anyone can download the model files from Hugging Face and run them on their own servers.
That open availability is a big deal. Most frontier AI models — the ones that perform at the highest level — are locked behind paid APIs from companies like OpenAI or Anthropic. DeepSeek V4 puts comparable capability into the hands of researchers, businesses, and developers worldwide, for free.
Why it matters
DeepSeek V4 matters for two reasons: what it can do, and what it costs.
On capability, V4-Pro claims to be the best open-source model for agentic coding — tasks where an AI works through a complex software problem over many steps, not just answering a single question. It also performs at a world-class level on math and STEM benchmarks, rivaling top closed-source models.
On cost, the open-weights release means you can self-host it. The API is also priced well below comparable Western models. DeepSeek has a track record here: its predecessor, V3, was priced at $0.27 per million input tokens — a fraction of what leading U.S. models charged at the time.
How it works (the basics)
DeepSeek V4 uses a design called Mixture of Experts (MoE). Here's a simple way to think about it: imagine a large team where only the relevant specialists show up for each job. V4-Pro has 1.6 trillion parameters in total, but only about 49 billion are "active" at any given moment. This makes it far cheaper to run than a model that uses all its capacity all the time.
Both V4-Pro and V4-Flash support a 1 million token context window — meaning they can read and reason over roughly 750,000 words in a single session. That's enough to hold an entire codebase or a stack of long documents in memory at once. This is enabled by a new architecture called DeepSeek Sparse Attention (DSA), which makes processing very long inputs more efficient.
The controversy: distillation allegations
DeepSeek V4's rise hasn't been without friction. Anthropic publicly accused DeepSeek (along with two other Chinese labs, Moonshot AI and MiniMax) of conducting large-scale distillation attacks — essentially, using millions of fake accounts to query Claude and harvest its responses as training data. Anthropic claims over 16 million such exchanges were generated through roughly 24,000 fraudulent accounts, in violation of its terms of service.
The concern isn't just about cheating. Anthropic argues that models trained this way may strip out the safety guardrails that were built into the original, creating AI that is powerful but less carefully controlled. The White House acknowledged the distillation threat in a memo framing it as a national security concern.
DeepSeek has not publicly responded to these specific allegations in the events covered here.
Recent developments
After the main V4 release, DeepSeek published DSpark — a speculative decoding module that makes V4 significantly faster in production. Speculative decoding is a technique where a smaller "draft" model predicts what the big model will say next, letting the system skip ahead and verify in bulk rather than generating one word at a time. DSpark achieves 57–85% faster per-user generation speeds and is released under an MIT license, meaning anyone can use it.
Researchers have also begun using V4 in benchmark studies — including safety evaluations and emotional-support chatbot assessments — reflecting how quickly it has become a standard reference point in the research community.
Where it fits in the landscape
DeepSeek V4 sits in a competitive field. On the Artificial Analysis Intelligence Index, it trails closed models like Claude Opus 4.8 (score 56) and GPT-5.5 (score 55). Among open-weights models, it competes with GLM-5.2 from Z.ai and Nvidia's Nemotron 3 Ultra, both of which also target the 1M-context, agentic-coding space.
What sets V4 apart is the combination: frontier-adjacent performance, open weights, a large and active community (V4-Pro has over 4.3 million Hugging Face downloads), and a lab that has consistently pushed inference efficiency further than its peers.
