In a fascinating experiment, developer Robert de Gallan deployed a swarm of five AI agents to deep-research the most promising stack for cheap, fully local AI-driven software engineering. After weeks of iterative runs, the agents converged on a specific pairing: the Qwen4 27B model combined with CoDoc, a proposed standard for AI-optimized documentation. The result challenges the prevailing narrative that local AI coding requires massive cloud infrastructure or frontier-class models to be viable.

The David and the Rock Strategy

The agents identified Qwen4 27B as the 'David' in this scenario. When quantized, the model fits comfortably on a single 24 GB consumer GPU. The analysis suggests that at this size, coding capability is sufficient for day-to-day engineering tasks, provided the agent has accurate context. The 'rock' is CoDoc, a standard where software libraries ship AI-optimized documentation directly inside the package. This includes a docs/README.md entry point and a docs/index.md catalog, all in plain markdown. Crucially, this documentation is version-matched by construction, offline, and deterministic, landing on disk alongside the code during installation.

Context Is the Real Bottleneck

According to the research agents, the primary failure mode for 27B local models is not a lack of raw intelligence but a lack of precise context. When a small model hallucinates a v5 API, it is a context problem, not a capability problem. CoDoc solves this by bundling the correct documentation with the library, eliminating the need for external retrieval systems. Evidence from the CoDoc white-paper supports this: Vercel’s internal evaluations on Next.js 16 APIs showed a 100% pass rate when using bundled docs with an AGENTS.md pointer, compared to just 53% without them. On-demand 'skill' variants maxed out at 79% and were not even invoked in 56% of cases.

Undermining the Cloud Moat

The most provocative finding is that CoDoc potentially collapses the primary moat of hyperscalers and big labs. These companies have long relied on paid context layers, including RAG pipelines, hosted doc intermediaries, and enterprise knowledge add-ons, to maintain relevance. By bundling context directly into the package via npm install or equivalent mechanisms, CoDoc removes the dependency on metered APIs. The standard originates from Amadeus, an unknown European IT company, which claims reductions in hallucinated APIs and deprecated patterns, though they admit to lacking rigorous public evidence beyond these statements. They plan to develop CoDocBench to address this gap.

Key Takeaways

  • Local AI coding viability hinges on context quality, not just model parameters.
  • Qwen4 27B is sufficient for most engineering tasks if paired with accurate, bundled documentation.
  • CoDoc threatens the business model of cloud-based context providers by making documentation offline and deterministic.
  • Amadeus, the creator of CoDoc, lacks first-hand public evidence but has no incentive to sell tokens, unlike cloud labs.

The Bottom Line

This isn't just about a better model; it's about a better architecture. If context can be shipped locally, the entire cloud dependency for coding agents evaporates, leaving hyperscalers with nothing to meter.