DeepSeek V4 ships a million-token context agents can use
Text2026-05-28
Source: huggingface.co
DeepSeek's open V4 release pairs a very large context window with strong reasoning at low cost, keeping open weights within reach of the frontier.
A million-token window is only useful if the model can actually attend across it, and that's the claim here: V4 is built so agents can keep whole codebases and document sets in context without the quality collapsing at the far end. Crucially, the weights are open.
That combination — frontier-ish capability, open licence, low serving cost — is why so much local tooling targets DeepSeek. You can run it through vLLM or llama.cpp; see the text & LLMs page for the stack, and hosting for somewhere to put it.