Flint: Local Coding Agents, Mac Memory and Running Cost

The BF16 release and the Mac package have different storage and hardware requirements. Recursive’s rec README lists M4 Pro, M5 Pro or M5 Max Macs, at least 48GB unified memory, about 30GB free disk and Node.js 22 for the pi harness. Those requirements describe the documented rec route; they do not establish compatibility with every Apple Silicon Mac.

One compatibility report illustrates why the chip matters as well as memory: rnella01 reported rec 0.1.0 stopping at its hardware check on an M4 Max with 48GB unified memory and macOS 26.6.1. In the same issue, ChrisCummins, identifying himself as an engineer at Recursive, said M4 Max had not been thoroughly tested. Check support for your chip and rec version before downloading the package.

The original Hub repository contains the BF16 checkpoint and an SGLang serving kit. Its roughly 79.6GB storage footprint belongs to that original release. The card documents one B300 or H200 GPU per replica for the supplied kit. Follow that kit’s own setup rather than applying the Mac package’s disk or memory figures to a BF16 server.

Follow the rec repository’s installation instructions on a supported Mac, then run rec doctor. The commands below open its preconfigured pi harness. First use downloads the model and harness. Try a bug fix with a known failing test, inspect the changes, and run the tests before accepting the result.

rec also supports OpenCode and an OpenAI-compatible local chat endpoint. rec serve prints readiness when the model is loaded; the documented base URL is http://127.0.0.1:11500/v1 with model flint. A client executes tool calls and returns their results. rec is optimized for a single session, and its server stays running after the harness closes; stop it when the job is finished.

The active count describes parameters used for one token’s computation. The model still contains all 39.8B parameters, and a working process also needs caches and runtime buffers. Quantization changes weight storage; it does not remove context memory or the operating system’s share of unified memory.

The card configures a 262,144-token window shared by prompt and generated output. That is a model context limit, not measured peak memory on a 48GB Mac. Start with the context your repository task needs, keep reasoning and tool results in the session budget, and measure before increasing it.

For a Mac you already own, measure wall power over loading, prompt processing, generation and tool execution. An illustrative 200W session lasting two hours at $0.20/kWh uses $0.08 of electricity: 0.2kW × 2h × $0.20. Add any allocated hardware cost and storage separately. These inputs describe a cost calculation, not a Flint performance result.

For a rented server, multiply the full billed allocation by your hourly rate, including loading, waiting and retries. Compare cost per accepted code change or completed task. A decode-only token rate leaves out prompt processing and the work performed by the agent’s tools.