DeepSeek V3.1 Officially Launched: Surpassing R1 in Performance, Faster and Free for All Users!

On August 21, 2025, DeepSeek announced the official release of its next-generation model DeepSeek-V3.1, now available on both its website and mobile app. Remarkably, this update is completely free for all C-end users.

Compared with its predecessor, DeepSeek-V3.1 delivers breakthrough improvements:

• Supports 128K ultra-long context, significantly expanding its ability to handle complex tasks;
• Introduces a hybrid inference architecture, allowing automatic switching between “thinking” and “non-thinking” modes;
• Much faster response time, generating accurate answers in a shorter period.

In terms of Agent (intelligent agent) capabilities, V3.1 has been further enhanced through Post-Training optimization, enabling smarter tool usage and multi-step task execution. Benchmark results show:

• In code repair (SWE), command-line complex tasks (Terminal-Bench), multi-step reasoning searches (browsecomp), and high-level interdisciplinary challenges (HLE), V3.1 significantly outperforms the previous R1-0528 model.
• The V3.1-Think mode, trained with compressed reasoning chains, reduces output tokens by 20%–50% while still maintaining equal or even better overall performance than R1-0528.

Regarding the much-anticipated DeepSeek R2, the company explained that its release has been slightly delayed due to challenges in training with domestically developed Ascend chips, as well as the team’s commitment to maintaining high-quality standards. Although rumors suggested an August release, the official statement has denied this. The launch is expected to happen in the coming weeks, with exact timing subject to further announcements.

With the full launch and free access of DeepSeek-V3.1, users can now experience faster responses, stronger reasoning, and smarter interactions—a major step paving the way for the arrival of R2.