Pro/deepseek-ai/DeepSeek-V3.1-Terminus

DeepSeek has released DeepSeek-V3.1-Terminus as a direct update to DeepSeek-V3.1, according to the company's official API documentation. The release notes explicitly name the variant and describe targeted improvements over the prior version: reduced Chinese/English language mixing and elimination of stray random characters, stronger Code Agent and Search Agent performance, and more stable, reliable outputs across benchmarks. DeepSeek confirms availability on App, Web, and API, and links the open-source weights at huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus, preserving clear creator attribution for the model rather than any serving provider. As a model-focused update, the announcement centers on user-reported pain points addressed at the model level rather than on hosting or routing details. DeepSeek frames V3.1-Terminus as a refinement that retains V3.1's core capabilities while tightening language consistency and agent reliability, both of which have direct product impact for coding and search-agent workflows built on the API. No quantitative benchmark deltas are included in the excerpt, so downstream readers should treat the claims as vendor-reported rather than independently verified, but the variant naming and weights link make this a primary reference for any DeepSeek-V3.1-Terminus work.

Artificial Analysis provides an independent third-party profile of DeepSeek V3.1 Terminus (Non-reasoning), explicitly naming the variant and documenting concrete technical specifications: 685B total parameters with 37B active per token, a 128k-token context window, text-in/text-out modality, MIT license, and a September 2025 release. The model scores 14 on Artificial Analysis' Intelligence Index, placing it above the median among comparable open-weight non-reasoning models, with listed pricing of $0.27 per 1M input tokens and $1.00 per 1M output tokens alongside a 50% cache discount. The page explicitly flags DeepSeek V3.1 Terminus as deprecated in favor of the newer DeepSeek V3.2 (Non-reasoning), and Artificial Analysis notes it only continues updating the default 10k-input benchmark workload while other metrics are historical. Speed, verbosity, and cost-per-intelligence figures are listed as N/A. Coverage is limited to the non-reasoning variant rather than the full hybrid reasoning/non-reasoning model, so this item is most useful as supporting technical reference rather than as a fresh release announcement.