CITIC Construction Investment published a research note on October 20 arguing that DeepSeek's Flash-model update, while keeping prices low, improves coding and automation-task performance and opens search and content-crawling capabilities; meanwhile Zhipu released a legal-specialized model and StepFun disclosed a long-horizon Agent training scheme. Improved model capability, call cost and tooling will expand Agents' reach in code development, data analysis and enterprise business systems, driving demand for enterprise workflows and inference.
On October 17, DeepSeek released V4.1-Flash-1017, synchronizing web, app and API with unchanged pricing. Building on September's V4.1 Flash, it improves reasoning, coding, automation and vision, and is optimized for frameworks such as DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. Through self-feedback reinforcement learning, the model better builds complete projects from requirements, demonstrated via 3D modeling, web games and Godot-based game development — extending from code generation to tool calling, result checking and multi-round correction.
Benchmark gains were clear: DeepSWE v1.1 rose from 74.2% to 77.0% (versus GPT-5.6 Sol 73.0% and Claude Opus 5 74.0%); Terminal-Bench 4.0 from 31.2% to 42.3% (GPT-5.6 Sol 39.9%); Automation-Bench from 54.8% to 60.5% (GPT-5.6 Sol 45.8%). The note recommends attention to enterprise-level AI applications, compute services, domestic chips and supernodes.
The research frames a shift from "parameter competition" to "calls, cost and scenario competition," favoring domestic compute, inference services and AI applications as model capability and tooling improve together.
Source: CITIC Construction Investment / Zhitong Finance
