DeepSeek-V3.1 notes: hybrid Think and Non-Think era

DeepSeek-V3.1, announced 21 August 2025, is the hybrid Think / Non-Think generation. One model, two modes, toggled on chat with DeepThink. Official notes: faster thinking than R1-0528, stronger agents, 128K context at that launch, Anthropic API format, strict function calling in beta. A later Terminus note focused on language consistency (fewer CN/EN mix-ups) and better code and search agents.

On the API of that era, deepseek-chat meant non-thinking and deepseek-reasoner meant thinking. Those aliases later pointed at V4-Flash during the retirement window. I mention that so old snippets do not surprise you.

Current work for me sits on V4-Flash and V4-Pro. V3.1 remains in this catalog because the official research list still names it.