Connect with us

AI

DeepSeek 3.1 AI model launched with agent skills

Published

11 months ago

on

DeepSeek

DeepSeek, China’s leading AI company, has launched DeepSeek version 3.1, its first model with agentic skills, faster thinking, as well as hybrid inference.

The company has confirmed that v3.1 offers Think and Non-Think modes within a single model, allowing flexible processing. It can think faster than the DeepSeek R1, providing a better performance for an improved user experience.

The version 3.1 Base is a result of continued pre-training with 840 billion tokens on the V3 foundation to extend long-context handling. Tokenizer and chat template have also received new updates. In addition to the main model launch, the company has also released open-source weights for both 3.1 Base and full 3.1.

Benchmark

This new model has shown superior results on SWE-Bench and Terminal-Bench with improved multi-step reasoning for intricate search operations, and gains in computational efficiency during thinking.

Here are the SWE and Terminal-Bench benchmarks

  • Swe-Bench Verified – 66.0 (DeepSeek 3.1), 45.4 (DeepSeek 3), 44.6 (DeepSeek R1)
  • SWE-Bench Multilingual – 54.5 (DeepSeek 3.1), 29.3 (DeepSeek 3), 30.5 (DeepSeek R1)
  • Terminal Bench – 31.3 (DeepSeek 3.1), 13.3 (DeepSeek 3), 5.7 (DeepSeek R1)

The new model has a lead over R1 on important benchmarks, including Browsecomp, HLE, xBench-DeepSearch, Frames, SimpleQA, and SealO.

DeepSeek 3.1 benchmarks

DeepSeek 3.1 benchmarks

Besides performance, DeepSeek 3.1 comes with massive output token efficiency on tasks such as AIME 2025, GPQA Diamond, and LiveCodeBench.

DeepSeek 3.1 output tokens efficiency

API and Pricing

DeepSeek has also updated its API; the deepseek-chat now stands for non-thinking mode, and deepseek-reasoner for thinking mode, both support a 128k context window. The API pricing for DeepSeek 3.1 remains unchanged, but the company has announced that the new pricing will start, and off-peak discounts will end on September 5, 2025.

The new input API will cost $0.07 per 1 million tokens (cache hit) and $0.56 per 1 million tokens (cache miss). The output API costs $1.68 per 1 million tokens.

(source)

Related Topics:

Mannoo specializes in Generative AI, Large Language Model (LLM), and Aerospace Science. Prior to delving into these fields, he was a Python programmer, a game designer, and an Android and iOS app developer with over 5 years of experience. He has prior writing experience in creative writing about smartphones and technology before working at Eonmsk.com. You can explore his X/TWitter and LinkedIn pages or contact him through his email.

Continue Reading