Recursive Unveils Flint: Compact 3-Billion Parameter On-Device Coding Model Trained by Autonomous Agent Sweeps

San Francisco artificial intelligence engineering venture Recursive unveiled Flint on Wednesday, October 21, introducing an ultra-compact 3-billion parameter coding model designed to run entirely locally on consumer laptops and edge developer workstations. Operating from the startup's research offices in San Francisco and Mountain View, company founders and department engineering officials demonstrated how the model delivers instantaneous, latency-free code completion and multi-file debugging without transmitting proprietary enterprise source code over public cloud networks.

Flint represents a major architectural milestone in synthetic data training, having been refined almost entirely through multi-agent reinforcement learning cycles where autonomous LLM supervisor agents generated, critiqued, and verified millions of intricate programming exercises and edge-case unit tests. On standard industry benchmarks including HumanEval and MultiPL-E, the 3B model achieved an impressive 78.4 percent pass rate, outperforming previous-generation 14-billion parameter open models while requiring less than four gigabytes of unified system memory.

Software engineers and enterprise privacy officers lauded the release for enabling developers working on sensitive military, healthcare, or corporate infrastructure projects to leverage modern AI coding assistants in completely air-gapped, zero-network environments. The model compiles natively to Apple Silicon Core ML and Qualcomm NPU runtimes, generating over 110 tokens per second while consuming minimal laptop battery power during extended development sessions.

Recursive released the pre-trained model weights, quantization kernels, and local VS Code extension plugins under a permissive open-access license on Hugging Face and GitHub. Software developers, engineering departments, and academic researchers can review technical documentation, benchmark telemetry, and installation instructions online at recursive.ai/flint. Technical inquiries regarding commercial enterprise support agreements may be directed to Recursive's San Francisco team.