A One-Year-Old Beijing Startup Just Beat GPT-5 and Claude on SWE-Bench. Investors Are Writing Checks Every Two Months.

On July 27, 2026, a one-year-old Beijing-based company called CiYuanWuXian announced its third funding round of the year. Founded in July 2025 by a former senior ByteDance executive, the company is now operating on a near-quarterly capital cadence that has investors lining up before the previous check clears. Angel round in January. Angel+ in June. Angel++ in late July. The cumulative raise is already in the hundreds of millions of RMB. The reason for the rush is not the team's pedigree. It is the SWE-Bench Verified score: 79.4% Pass@1, posted in December 2025, ahead of GPT-5, Claude, and every other frontier model on the public leaderboard at the time. The company's product, an enterprise AI coding agent called InfCode, also took the #1 spot on Multi-SWE-bench's Java track the same month. In a market where Cursor and Cognition have already raised hundreds of millions at billion-dollar valuations, a one-year-old Chinese competitor is forcing the question of whether the enterprise coding-agent stack is going to look very different by 2027.

The founder, who spent six years leading monetization product at ByteDance before starting CiYuanWuXian, has been clear about positioning from day one: InfCode is not a Cursor competitor. Cursor and Windsurf sell subscriptions to individual developers. InfCode sells to enterprise platform teams that need an AI agent that can read a million-line legacy Java codebase, pass internal security review, and ship to production. That is a different sale. The contract size is millions, not $20 a month. The sales cycle is six months, not a self-serve signup.

The Benchmark Story

SWE-Bench Verified is the closest thing the industry has to a standardized test for autonomous code generation. It scores whether a model can resolve real GitHub issues end-to-end. In December 2025, InfCode hit 79.4% Pass@1, leading the public leaderboard for several weeks. For comparison, GPT-5 sat in the mid-70s, Claude Opus 4.x in the high-70s. The same period saw InfCode take the Java #1 on Multi-SWE-bench, a harder variant that tests cross-file refactoring in real enterprise repositories.

Two things make the result unusual. First, InfCode is a relatively compact model — the company has not publicly disclosed parameter count, but the inference cost is reportedly a fraction of GPT-5.6 Sol. Second, the training data leans heavily on enterprise Java and Python codebases, not the GitHub-scrape data most open-source coding models use. That domain specificity is the entire point: the team is not chasing general coding ability. They are chasing the specific failure modes that show up when a junior engineer with no codebase context gets dropped into a 10-year-old Spring Boot monolith.

Why Three Rounds in Twelve Months

The funding cadence is itself a signal. The Angel round came from a software-focused CVC. The Angel+ in June was led by Huakong Fund and Shuimu Venture Capital, with Xiamen Innovation Capital following. The Angel++ on July 27 was led by Linxin Investment with Huakong Fund continuing. The investors are not tourists. They are software-industry specialists who have watched Cursor, Cognition, and Decagon compress from garage project to $1B+ revenue in 18 months, and they are not going to miss the same curve in China.

The capital is being deployed against a very specific roadmap. The company has stated that the funds will be used to build what they call "China's first fully autonomous, high-security, self-evolving enterprise AI Agent infrastructure platform." The partnership with Tsinghua University and Beihang University is not academic window-dressing — it is an explicit talent pipeline. The competitive moat is not the model. The competitive moat is the proprietary corpus of enterprise code, the fine-tuning infrastructure, and the security review tooling that gets the model through a Fortune 500 CISO's office.

Key Takeaways

• CiYuanWuXian, founded July 2025 by a former ByteDance executive, has closed three funding rounds in twelve months: January (Angel), June (Angel+), and July 27 (Angel++).
• InfCode hit 79.4% Pass@1 on SWE-Bench Verified in December 2025, ahead of GPT-5 and Claude at the time, and took Java #1 on Multi-SWE-bench.
• The product targets enterprise platform teams, not individual developers — multi-million RMB contracts versus $20/month subscriptions.
• Strategic partners include Tsinghua University and Beihang University for talent pipeline and joint research.
• Investors include Huakong Fund, Shuimu Venture Capital, Linxin Investment, and Xiamen Innovation Capital — all software-specialist funds.

My Take

The most interesting thing about CiYuanWuXian is not the benchmark. The benchmark is the door-opener. The interesting thing is the company has decided that the Cursor model — a polished IDE for individual developers — is not the highest-margin game in AI coding. The higher-margin game is selling the agent to a CIO who needs to ship AI into an existing Java backend without breaking compliance, audit, or SLA. That customer does not want a $20 subscription. They want a $5M annual contract with a dedicated engineering team, an on-prem option, and a 99.9% uptime guarantee. The fact that InfCode is already benchmark-competitive with GPT-5 while being trained on enterprise-specific data is what makes the round look cheap at hundreds of millions. If Cursor is the consumer opportunity in AI coding, InfCode is the enterprise opportunity — and the enterprise market is at least 10x larger. The risk is execution: enterprise sales cycles are brutal, and a one-year-old company has to scale account management, security, and support faster than its hiring. The capital gives them a runway. It does not give them a sales pipeline.

Frequently Asked Questions

What is InfCode and how is it different from Cursor?

InfCode is an enterprise AI coding agent built by Beijing startup CiYuanWuXian. Cursor targets individual developers with a $20/month IDE subscription. InfCode targets enterprise platform teams with multi-million RMB annual contracts, on-prem deployment, and full integration into existing security and compliance review processes. The two products compete in adjacent markets, not the same one.

How does InfCode score higher than GPT-5 on SWE-Bench?

InfCode was trained on a domain-specific corpus heavy in enterprise Java and Python codebases, rather than the generic GitHub scrape used by most open-source coding models. The team has optimized for the specific failure modes that occur in legacy enterprise repositories — large monorepos, cross-file refactoring, framework-specific conventions. That focus shows up in the 79.4% Pass@1 score on SWE-Bench Verified and the Java #1 on Multi-SWE-bench in December 2025.

Who are the investors in CiYuanWuXian?

The Angel round came from a software-focused CVC. The Angel+ in June 2026 was led by Huakong Fund and Shuimu Venture Capital, with Xiamen Innovation Capital. The Angel++ round announced July 27, 2026, was led by Linxin Investment with Huakong Fund continuing. All three rounds close within twelve months of founding.

Can international developers use InfCode?

InfCode is currently focused on the Chinese enterprise market and is not publicly available outside mainland China. The company has not announced an international launch timeline. For developers outside China, the equivalent tools in the current market are Cursor, Cognition's Devin, and GitHub Copilot Workspace — see the aifreetool.site AI programming tools directory at aifreetool.site/tool-category/ai-programming-development/ for a broader comparison.