The Zhitong Finance App learned that since October, Fangzheng Securities released a research report saying that since October, marginal changes in major domestic models include: DeepSeek synchronously updates models, tool APIs, and weights; Kimi expands training and completes high-speed versions; MiniMax enhances agent capabilities and reduces task costs. The computing power industry chain model parameters and inference demand jumped at the same time, and the computing power demand for training and reasoning expanded rigidly over a long period of time. The domestic model opened the ceiling of global demand, and commercialization accelerated. Large-scale agent deployment and data flow transfer are driving security protection requirements.
Fangzheng Securities's main views are as follows:
I. DeepSeek V4.1-Flash-1017: Simultaneous updates of models, tools and weights
The model was launched on October 17 on the website, app, and API. Officially, the new version achieves Pareto improvements in performance and cost, completing tasks faster, and improving end-to-end project development capabilities through self-feedback; the official evaluation scores for DeepSWE v1.1 and Automation-Bench v1.0.6 were 77.0 and 60.5, respectively. Both Search and Fetch API calls are priced at $0.75/1000; related server-side tool calls in DeepSeek Harness, Claude Code, and Codex are scheduled to be billed from 16:00 UTC on October 31; model weights have been released in Hugging Face and ModelScope.
2. Kimi K3.1 Pro: Expanded training, complementing the high-speed version
Kimi released K3.1 Pro and Highspeed on October 24, positioned as a 3T model, supporting 1 million token contexts; Highspeed's output speed is 2 to 3 times that of the standard version, enhancing coding, 3D creation, financial research, and long-link agent capabilities. After expanding on K3, K3.1 Pro trains computing power and carries out joint reinforcement learning in hundreds of downstream fields. The standard API input and output prices are 13 yuan and 68 yuan/million tokens, respectively, and Highspeed is 26 yuan and 136 yuan/million tokens respectively; the full weight is scheduled to be released before November 4, and has not been released as of October 26.
3. MiniMax M3.1 Flash: Agent efficiency improvement and open weight
The model was released on October 19 to upgrade Agentic tasks, programming, computer use, and network security capabilities; the official disclosure revealed that the speed of online inference was increased by nearly 3 times, the comprehensive use cost of Agentic tasks was reduced by 70% compared to the previous generation, the output speed for everyday use scenarios can reach 200+ TPS, and the P50 is 170+ TPS. The API input, output, and cache reading prices are 1 yuan, 4 yuan, and 0.1 yuan/million tokens, respectively. On October 23, the model weight was opened, with total parameters 428B and activation parameters 23B; within 72 hours after launch, real users processed more than 100 trillion tokens.
Risk warning: model iteration and commercialization falling short of expectations; computing power, production capacity and supply chain risks; regulatory policy changes; macro and geopolitical risks.