Build Your Own Cursor This Weekend. Yes, the One SpaceX Just Paid $60 Billion For.

2026-06-21 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, long

Summary

Cursor's in-house coding model, Composer 2, acquired by SpaceX for \$60 billion in June 2026, was built starting from Moonshot AI's open-weight Kimi K2.5 checkpoint. This demonstrates that a "frontier-level" coding assistant can be constructed primarily through integrating open-source components: a Visual Studio Code fork for the editor, a local inference server, and open-weight models. The core architecture involves a two-model setup: a small, fast Fill-in-the-Middle (FIM) capable model like Mistral's Codestral (22 billion parameters, ~95% FIM accuracy) for real-time autocomplete, and a larger model such as Qwen3-Coder-30B (30 billion parameters, fitting in ~19GB on a 24GB GPU) for chat and agentic tasks. This approach enables developers to build a functional, local AI coding assistant, processing tokens on their own hardware without external servers.

Key takeaway

For AI Engineers or ML Engineers building internal coding tools, you can now deploy a powerful, local AI assistant without proprietary models. Utilize open-source components like VS Code and Continue.dev, pairing a FIM-trained model (e.g., Codestral) for autocomplete with a larger model (e.g., Qwen3-Coder-30B) for chat. This approach keeps your code on-premises and eliminates per-token costs, offering a robust alternative to cloud-based solutions.

Key insights

A frontier coding assistant can be built by integrating open-source components and open-weight models, utilizing a two-model architecture.

Principles

• Start with open-weight model checkpoints.
• Separate models for autocomplete and chat.
• Aggressively gather relevant code context.

Method

Build a coding assistant by integrating VS Code with a Continue.dev extension, serving a FIM-trained autocomplete model (e.g., Codestral via Ollama) and a larger chat/agent model (e.g., Qwen3-Coder-30B via vLLM) locally.

In practice

• Deploy Codestral for FIM autocomplete.
• Use Qwen3-Coder-30B on 24GB GPUs.
• Serve models with Ollama or vLLM.

Topics

• AI Coding Assistants
• Open-weight LLMs
• Local LLM Deployment
• Fill-in-the-Middle
• VS Code Extensions
• Model Quantization

Code references

• vllm-project/vllm

Best for: AI Engineer, Machine Learning Engineer, Software Engineer

Related on AIssential

• 5 Cool Things I Did with Local Language Models — Local LLMs offer superior privacy, cost, and offline capabilities for sensitive data and proprietary workflows.
• 😺 Cursor’s $60B SpaceX deal is official — The AI coding sector is consolidating, with major players investing in advanced models and workflow integration.
• OpenCode Supercharged: BEST AI Coding Agent Ever + FULLY FREE! BYE Gemini CLI & ClaudeCode! — Oh My Open Code transforms basic AI coding agents into robust, multi-agent development teams with advanced orchestration.
• Another try at LLM workflow automation — Effective LLM automation requires extensive scaffolding and iterative tooling, not just smarter models.
• Stop Paying for API Keys: How to Build with Free and Fast LLMs — Groq offers a cost-free, fast platform for building AI prototypes with open-source LLMs like Llama and Qwen, bypassing API costs and hardware needs.
• 5x for Free : The Local Coding Stack — Local AI models, particularly MoE architectures, now offer competitive performance for coding with privacy and cost benefits.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.