DeepSeek-V3.1

DeepSeek announced DeepSeek-V3.1 on August 20, 2025, positioning it as the company's first step toward the agent era. The model introduces a hybrid inference design that lets a single model operate in both Think and Non-Think modes, with DeepSeek-V3.1-Think reportedly reaching answers faster than DeepSeek-R1-0528. Post-training improvements target stronger tool use and multi-step agent task performance, with gains cited on SWE and Terminal-Bench benchmarks. The release ships open-source weights for both V3.1 Base and V3.1 on Hugging Face, built on a V3.1-Base checkpoint that continued pretraining with 840 billion tokens for long-context extension. The tokenizer and chat template were updated, and the context window expands to 128K. API changes map deepseek-chat to non-thinking mode and deepseek-reasoner to thinking mode, add Anthropic API format support, and enable strict Function Calling as a beta feature.

DeepSeek's API documentation page confirms the DeepSeek-V3.1 release as the company's first step toward the agent era, introducing hybrid Think and Non-Think inference in a single model. DeepSeek-V3.1-Think is described as reaching answers faster than DeepSeek-R1-0528, with stronger agent skills from post-training that boost tool use and multi-step task execution. Results on SWE and Terminal-Bench are highlighted as improved. The API update routes deepseek-chat to non-thinking mode and deepseek-reasoner to thinking mode, both supporting a 128K context window. Anthropic API format compatibility and a strict Function Calling beta are added, while the V3.1 Base checkpoint adds 840 billion tokens of continued pretraining over V3 for long-context extension. Open-source weights for V3.1 Base and V3.1 are published on Hugging Face alongside a new tokenizer configuration.

RunPod published a technical analysis examining DeepSeek V3.1's architectural changes relative to the V3-0324 model, describing V3.1 as a breakthrough hybrid reasoning model that dynamically toggles between fast inference and deep chain-of-thought logic within a single architecture. The piece situates the V3.1 release The analysis details how V3.1's hybrid behavior is governed by chat template parameters rather than separate model paths: the thinking mode uses the

DeepSeek V3.1 is described as a hybrid follow-on to DeepSeek-V3 that unifies chat, reasoning, and coding in a single model. It supports a roughly 128k-token context window and precision formats from BF16 down to FP8 for efficient inference on modern accelerators. The release is positioned as a single model spanning these capabilities rather than separate domain-specialized variants. On the Aider coding benchmark, DeepSeek V3.1 reportedly scores 71.6%, indicating top-tier code capability among open models. Community analysis has surfaced special tokens hinting at built-in "thinking" and search hooks, though official documentation of those mechanisms is not yet posted. Hugging Face already shows a base entry for "DeepSeek-V3.1-Base," indicating the family is being published in multiple variants as quantized and fine-tuned derivatives appear.

DeepSeek-V3.1 is a large hybrid reasoning model (671B parameters, 37B active) that supports both thinking and non-thinking modes via prompt templates. It...