DeepSeek V3.1-Terminus is an iterative refinement of DeepSeek's V3 language model family, built to sharpen performance on tasks where the original V3.1 showed user-reported weaknesses. The model ships with hybrid thinking capabilities, letting developers toggle between reasoning and non-reasoning modes depending on the task at hand, and it is designed to function as a strong foundation for AI agents—particularly those built around code generation and web search workflows. Its architecture inherits the Mixture-of-Experts backbone and multi-head latent attention mechanisms that define the broader DeepSeek V3 series, while the updated Terminus checkpoint addresses language consistency issues such as unwanted Chinese-English mixing and stray abnormal characters that users flagged in earlier releases.

The Terminus update was driven by community feedback and refined through targeted optimizations across agentic benchmarks. Benchmark results show meaningful gains: Humanity's Last Exam scores climbed from 15.9 to 21.7, and BrowseComp—an evaluation of tool-augmented agent performance—jumped from 30.0 to 38.5, signaling a substantial uplift in search agent reliability. Code-related benchmarks such as LiveCodeBench held steady near 74.9, while general reasoning metrics like MMLU-Pro and GPQA-Diamond showed modest but consistent improvements. Released under an MIT license with open-source weights available on Hugging Face, the model invites developers to inspect, fine-tune, and deploy it directly, extending the openness that has become a hallmark of the DeepSeek family. SambaNova's deployment reports throughput exceeding 200 tokens per second, positioning V3.1-Terminus among the fastest open-weights reasoning models available for production agent pipelines.