Open-Source AI

Today's Stories

1. 1

   DoorDash, startups turn to cheaper Chinese AI models as U.S. rivals get pricier

   DoorDash is launching an experimental tool that uses Moonshot AI's model, while other startups like Cursor and Lindy have adopted Chinese AI models from Moonshot, DeepSeek, and others to reduce costs. Airbnb and Siemens are also experimenting with Chinese AI providers including Alibaba and DeepSeek. U.S. AI companies like OpenAI, Google, and Anthropic offer advanced models but at higher costs. As token and usage fees rise, companies are drawn to cheaper Chinese open-source alternatives, especially when they can run models locally to keep proprietary data in-house rather than sending it to outside providers. A March 16, 2026 study from Hugging Face found that Chinese open-source models accounted for 41% of downloads.

   Security experts warn that adopting Chinese models risks "data sovereignty violations" and "exposure [of] proprietary code and user data to foreign surveillance," though some analysts suggest companies may blend models—using Chinese AI for certain tasks and U.S. providers like Anthropic for others rather than a wholesale switch.

2. 2

   Capital One open-sources VulnHunter, AI tool that maps code flaws like attackers would

   Capital One released VulnHunter on Thursday, an open-source AI security tool available on GitHub under an Apache 2.0 license. The tool scans source code for exploitable vulnerabilities, maps how an attacker would reach them, and proposes targeted fixes before code ships to production. Capital One, still known for a 2019 data breach that compromised personal information of roughly 106 million people across the United States and Canada and cost the bank an $80 million（約130億円） federal fine, is now contributing offensive AI capabilities as a public defensive resource—a shift in how the company manages security risk.

   VulnHunter uses what Capital One calls an 'attacker-first forward analysis' workflow, beginning at the points where a real adversary would enter the system, which represents an ambitious approach to vulnerability detection for a major financial institution.

3. 3

   Kimi K3 exposes frontier labs' dilemma: keep AI models secret or lose them to China

   Moonshot, a Chinese startup, released Kimi K3, an open-weight AI model that performs at or near the frontier level of US labs like Anthropic and OpenAI. The release sparked market concern about China closing the technological gap, though running the most capable 2.8-trillion-parameter version requires a cluster of Nvidia GPUs costing several million dollars. Kimi K3 fits an established pattern where Chinese firms allegedly extract training data from American frontier models and use it to train open-source models anyone can download. This model is unsustainable for US labs, forcing them to choose between moving faster (constrained by US government security vetting that delays releases by a month) or abandoning the public model release strategy entirely.

   Frontier labs face a strategic fork: keep models secret to prevent distillation and avoid security delays, which would turn them into holding companies with massive advantages over other businesses worldwide—a future that open-source advocates and big-tech critics say they want to avoid.

4. 4

   Brex builds AI agent control layer at network level, not in rules

   Brex created CrabTrap, an open-source HTTP/HTTPS proxy that intercepts all network traffic from AI agents, examines policy rules, and uses an LLM-as-a-judge to approve or deny requests. The company found that traditional guardrails could not contain what agents were doing with real credentials like API keys and OAuth tokens. Brex's approach addresses a gap in how AI agents are currently governed—frameworks like OpenClaw enable agents to act, but lack enterprise-scale safeguards. By enforcing policy at the network layer rather than in the agent's code, organizations can audit and control agent behavior in real time, even when agents have genuine credentials to systems that matter.

   Brex CEO Pedro Franceschi frames this as a shift in how IT leaders should think about agent governance: moving from SDK-level permissions and model guardrails to centralized network control. How widely CrabTrap is adopted, and whether other enterprises adopt similar network-layer enforcement, will signal whether this architectural approach becomes standard practice.

5. 5

   Xi positions China as leader of open-source AI challenge to US

   Chinese leader Xi Jinping made his first appearance at the World Artificial Intelligence Conference, framing Beijing as the leader of a new global AI order and calling on countries to seize the "historic opportunity" of open-source models. At the same event, Chinese tech firm Moonshot unveiled what it claimed is the world's biggest open AI model. Xi's framing of open-source AI as a solution to "new historical injustices" emerging from unequal access to the technology signals Beijing's intent to position itself as an alternative to US-led AI development. Though US systems remain the most advanced, experts believe China's leadership in other layers of the AI stack — including abundant access to chips and energy — could give its champions an advantage in competing for global influence.

   The practical impact of China's open-source strategy on the global AI competitive landscape, particularly whether it shifts how developing nations adopt and build AI systems.

6. 6

   Robbyant releases LingBot-VLA 2.0 universal AI brain for robots

   Robbyant, an embodied AI company within Ant Group, has released LingBot-VLA 2.0, an upgraded vision-language-action (VLA) model trained on 60,000 hours of real-world physical data from 20 robot morphologies across 17 manufacturers. The model expands support for head, waist, end-effectors, and mobile chassis control, and improves inference efficiency by 3 times compared to the previous generation while keeping latency under 150 milliseconds. The embodied AI industry has lacked a truly universal brain for industrial-scale robot deployment. LingBot-VLA 2.0 addresses this bottleneck by demonstrating superior cross-morphology generalization—on the Shanghai Jiao Tong University's GM-100 benchmark, it outperformed both π0.5 and GR00T N1.7 on dual-arm manipulation, and surpassed π0.5 on long-horizon mobile manipulation tasks. The 3× inference improvement and sub-150-millisecond latency significantly lower the barrier for real-time commercial applications.

   Robbyant is conducting comprehensive commercial pilot testing with hardware partners Leju and Ti5Robot, and enterprise customers GuoDa Drugstore and Longsheng Technology in retail sorting, logistics, and industrial environments. The company is also partnering with GenRobot.ai to build standardized data ecosystems.