Cursor Built Its New Coding AI on Chinese Open-Source Models, Users Discover

Cursor unveiled Composer 2 on Saturday, pitching it as a faster, cheaper alternative to generalist AI models for software development. Within hours, a detail buried in Forbes' coverage stopped developers mid-scroll: the model was reportedly built on Chinese open-source foundations including DeepSeek, Kimi and Qwen, then fine-tuned using Cursor's proprietary data.

For enterprise customers already navigating heightened scrutiny over AI supply chains, the disclosure landed with weight. Forbes reported the upstream models directly, though other outlets covering the launch did not enumerate the same base models. Cursor has not publicly confirmed which specific model versions were used, and the exact version of Kimi, whether it is K2.5 or another release, has not been verified independently.

The model itself is built around a narrow mandate. Cursor cofounder and research lead Aman Sanger was direct about the scope: "It won't help you do your taxes. It won't be able to write poems." Training focused exclusively on coding-related data, a deliberate constraint Cursor argues lets it build a smaller, cheaper model without sacrificing performance on development tasks.

Roughly 20 AI researchers worked on the Composer series, according to Forbes. The training pipeline combined continued pretraining with scaled reinforcement learning, oriented toward long-horizon tasks that require hundreds of sequential actions: reading a repository, deciding what to edit, modifying multiple files, running tests, interpreting failures and iterating toward a working outcome. Custom tooling built for training included codebase-wide semantic search and structured feedback loops to improve the model's grasp of complex, multi-file projects.

Composer 2 ships in two pricing tiers. Composer 2 Standard costs $0.50 per million input tokens and $2.50 per million output tokens. Composer 2 Fast, which Cursor made the default experience, runs $1.50 and $7.50 respectively. VentureBeat, which reported the pricing, also noted the comparative benchmark positioning: Composer 2 outperforms Claude Opus 4.6 but trails GPT-5.4, according to that outlet's coverage. The specific benchmarks and evaluation methodology behind those claims have not been independently confirmed.

Cursor's broader economics reflect the pressure any AI intermediary faces when the underlying model providers are also shipping competing products. Forbes reported that consumer subscriptions currently run at negative margins, while business plans, including a Teams offering aimed at startups and enterprise contracts for larger organizations, operate at positive margins. The company's predecessor model, Composer 1.5, was described by Forbes as the second-most popular model on the platform and meaningfully cheaper to run than licensing Anthropic's larger models.

Developer reaction split along predictable lines. AI engineer Alex Havryleshko praised the context capabilities, writing that Composer "gives it power with awareness and project-level intelligence." Product designer Alex Nucci offered a cooler read: "Biggest problem with cursor is it's too agreeable. let's see how this does."

VentureBeat noted that some power users described moving to Anthropic's Claude Code, citing preferences for terminal-first workflows and longer-running agent behavior, though the outlet cautioned those comments are unverified and may not reflect broader user sentiment.

Internally, Cursor is already pushing further. Forbes reported the company is experimenting with coordinating hundreds of simultaneous agents, a mode internally called "grind mode," though engineers have observed that agents, when working alongside many simulated coworkers, can become less productive, a dynamic Cursor researchers liken to human team behavior.

Cursor, backed by parent company Anysphere at a reported valuation of $29.3 billion according to VentureBeat, launched its first AI coding assistant in 2023. Composer 2 represents its most direct argument yet that specialization beats scale.