Summary
On August 14, 2026 Cloudflare added DeepSeek V4 Flash and DeepSeek V4 Pro to Workers AI, the first models on the platform with a full 1,048,576-token context window. Both support reasoning (thinking mode) and multi-turn function calling, positioning them for long-horizon agent workloads rather than single-turn chat.
What changed
Cloudflare made DeepSeek V4 Pro and the faster, lower-cost V4 Flash available on Workers AI via the Workers AI binding, REST API, OpenAI-compatible endpoint, and AI Gateway. Both models ship a 1,048,576-token context window, thinking mode, and function calling; access requires a Workers Paid plan or prepaid AI Gateway credits.
Why it matters
A million-token context on an edge inference platform matters for agents specifically because multi-step runs accumulate history fast, and function calling plus reasoning let developers build tool-using agents that keep large codebases or long tool traces in context. It moves Cloudflare's serverless AI from chat-scale to agent-scale workloads at the network edge.
Evidence excerpt
The first Workers AI models with a full one million (1,048,576) token context window. Both models support reasoning, function calling, and long context; V4 Flash is the faster, lower-cost sibling.