Last week, a developer on the LocalLLaMA subreddit ran Qwen's 9B parameter model through Aider's coding benchmark and scored 19.1%. Then, without changing a single weight, they swapped the agent scaffold and scored 45.6%. Same model. Same hardware. Same benchmark. The only variable was the software wrapping around the model: shorter system prompts, tighter tool schemas, single-step edits instead of multi-turn planning chains. The scaffold designed for small local models outperformed the one designed for frontier models by a factor of 2.4.
This is not a story about one benchmark result. It is a story about where value accrues in AI systems, and the answer is shifting faster than most people realize.
For the past two years, the dominant assumption has been that model capability is the binding constraint. Pick the smartest model you can afford, feed it your problem, and the quality of the output tracks the quality of the model. This assumption drove a rational strategy: wait for the next model release, upgrade, and enjoy the improvement. It also concentrated leverage with the model providers. If the model is the bottleneck, then OpenAI, Anthropic, and Google hold the cards.
The scaffold result inverts that logic. When a 9B model with the right framing outperforms itself by 140%, the implication is that we have been leaving enormous performance on the table not because our models are too small, but because our orchestration is too generic. The model is a fixed cost. The scaffold is the variable that moves the needle.
This pattern has a historical precedent worth studying. In the early cloud era, roughly 2008 to 2012, the conventional wisdom was that compute was the scarce resource. Buy more servers, get better performance. What actually happened was that orchestration ate the value chain. AWS did not win because it had better servers than IBM. It won because it wrapped commodity hardware in a scaffold of APIs, auto-scaling, and managed services that made the hardware dramatically more useful per dollar. The companies that understood this early built on AWS. The ones that kept buying bigger iron fell behind.
The same inversion is playing out now in AI tooling, and the evidence is converging from multiple directions. Matt Webb published a thesis last week arguing that headless services will be the winning pattern as AI agents become the primary interface layer. Strip the UI. Expose the API. Let the agent be the scaffold. Webb's argument is downstream of the same insight: the orchestration layer, not the capability layer, is where differentiation will live.
Meanwhile, the local inference stack is getting fast enough to make scaffold optimization practical at every scale. Speculative checkpointing just merged into llama.cpp's mainline, enabling the runtime to save draft model state and resume on rejection rather than recomputing from scratch. Users are reporting speed increases of up to 665% on code editing tasks when combined with ngram-map speculative decoding. That is not a model improvement. That is pure scaffold engineering, optimizing how tokens flow through inference rather than changing what produces them.
Zoom out further and the convergence becomes hard to ignore. Qwen 3.6, the model dominating local LLM discussion for a fourth consecutive day, is notable not because it is the largest or most capable model available. It is notable because its efficiency profile, 35 billion parameters with only 3 billion active, makes it responsive to exactly this kind of scaffold optimization. People are not just benchmarking it. They are switching their daily workflows to it, which only happens when the surrounding tooling makes a model feel faster and more reliable than its raw capabilities would suggest.
The strategic implication is uncomfortable for anyone whose plan is to wait for better models. If the scaffold is the product, then competitive advantage comes from how you frame, constrain, and orchestrate AI, not from which provider's API key you paste into your config file. Two teams using the same model with different scaffolds will get meaningfully different results. And unlike model training, scaffold design is something any team can iterate on today, without a GPU cluster or a research lab.
This also reframes the build-versus-buy decision for AI tooling. Most coding agent frameworks, Aider, Claude Code, and their peers, are optimized for frontier-class models because that is where the highest absolute performance lies. But optimization for the frontier is not optimization for your specific context. The developer who got 45.6% out of a 9B model did it by stripping away the patterns that frontier-optimized scaffolds rely on: long chain-of-thought prompts, multi-turn planning, verbose tool descriptions. These patterns help large models think. They confuse small ones. The right scaffold is context-dependent, and context is something you know better than any framework author.
The next six months will test this thesis. As Qwen, Llama, and other open-weight models continue to improve at the 7B to 35B scale, the teams that invest in scaffold engineering will pull ahead of those relying on raw model upgrades alone. The gap between a well-scaffolded local model and a poorly-scaffolded cloud model is already narrowing. When it closes, the value of model access as a competitive moat drops to near zero. What remains is the quality of the frame you build around it.
Watch for the tooling. The frameworks that let you tune prompt structure, tool schemas, and edit patterns per model size, rather than shipping one-size-fits-all defaults, will be the ones that matter. The model is the engine. The scaffold is the car. Nobody buys an engine.