Founder: Competition among domestic large models shifts toward task delivery, with simultaneous leaps in Agent capabilities and inference efficiency.

Founder released a research report stating that since October, the marginal changes in major domestic large models mainly include: DeepSeek simultaneously updated its model, tool API, and weights; Kimi expanded post-training and added a high-speed version; MiniMax improved Agent capabilities and reduced task costs. Model parameters and inference demand in the computing power industry chain are rising simultaneously, with long-term rigid expansion in training and inference computing power demand. Domestic models are opening up the global demand ceiling, accelerating commercialization. Large-scale Agent deployment and data flow are driving up demand for security protection. Founder's main points are as follows: 1. DeepSeek V4.1-Flash-1017: Simultaneous updates to model, tools, and weights The model launched on the web, App, and API on October 17. The company stated that the new version achieves a Pareto improvement in performance and cost, with faster task completion and shorter outputs, and enhances end-to-end project development capabilities through self-feedback reinforcement learning; official evaluations show DeepSWE v1.1 and Automation-Bench v1.0.6 scores of 77.0 and 60.5, respectively. Search and Fetch API call pricing is US$0.75 per thousand calls; related server-side tool calls in DeepSeek Harness, Claude Code, and Codex are scheduled to be billed starting at 16:00 UTC on October 31; model weights have been released on Hugging Face and ModelScope. 2. Kimi K3.1 Pro: Expanded post-training, added high-speed version Kimi released K3.1 Pro and Highspeed on October 24, positioned as a 3T-class model supporting a 1-million-token context; Highspeed output speed is 23 times that of the standard version, strengthening Coding, 3D creation, financial research, and long-chain Agent capabilities. K3.1 Pro expands post-training compute on the basis of K3 and conducts joint reinforcement learning across hundreds of downstream domains. Standard version API input and output prices are RMB13 and RMB68 per million tokens, respectively, while Highspeed is RMB26 and RMB136 per million tokens; full weights are scheduled for release before November 4 and had not been released as of October 26. 3. MiniMax M3.1 Flash: Agent efficiency gains and open weights The model was released on October 19, upgrading Agentic tasks, programming, Computer Use, and cybersecurity capabilities; the company disclosed that online inference speed increased by nearly 3 times, comprehensive usage cost for Agentic tasks decreased by 70% compared with the previous generation, and output speed in daily usage scenarios can reach 200+ TPS, with P50 at 170+ TPS. Its API input, output, and cache read prices are RMB1, RMB4, and RMB0.1 per million tokens, respectively. Model weights were opened on October 23, with total parameters of 428B and activated parameters of 23B; within 72 hours of launch, real users processed more than 100 trillion tokens. Risk warnings: Model iteration and commercialization falling short of expectations; computing power capacity and supply chain risks; changes in regulatory policies; macroeconomic and geopolitical risks.