Overview

• DeepSeek has rolled out V3.1‑Terminus across its app, web client, mini‑program and API, replacing previous V3.1 endpoints.
• Terminus exposes two modes—a non‑thinking chat model and a thinking reasoner—each with a 128k context window and defined output limits (chat 4k default/8k max; reasoner 32k default/64k max).
• The company targets language consistency and abnormal‑character failures, and journalists could not reproduce the prior “极” output bug or multilingual mixing in new API tests.
• Official comparisons show 0.2%–36.5% improvements on non‑agent benchmarks with a pronounced HLE uptick, alongside minor declines on select tests.
• Hands‑on trials report stronger Code and Search Agent behavior, while the model is now open‑source on Hugging Face and ModelScope with pricing set at 0.5 RMB per million input tokens on cache hit, 4 RMB on miss, and 12 RMB per million output tokens.