DeepSeek V3.2

DeepSeek V3.2 is an older open-weight model from DeepSeek, and it suits anyone who wants very low token prices for reasoning and tool use under a plain MIT licence. The Hugging Face card lists 685B parameters, and says the weights and code are MIT licensed. It introduces DeepSeek Sparse Attention, aimed at cutting the cost of long inputs. The card gives no context length or exact date. OpenRouter lists it from 1 December 2025 with a 163,840-token window, text only, at $0.26 per million input tokens and $0.38 per million output tokens. The card warns that its Speciale variant has no tool calling, that no chat template is shipped, and that the output parser is not production ready without extra error handling. Its benchmark figures are the maker's own. Figures come from the Hugging Face card and the OpenRouter model list. We have not run it ourselves. We would use it as a budget model for batch work, and move to the V4 line for anything new.