Last week, on Reddit, a user shared their experience with a local language model: Qwen 3.6 35B A3B. Their post, titled “local already feels good enough,” describes a focused use case – coding, technical planning, and hardware setup – and delivers a clear verdict: the LLM “hasn't skipped a beat if it has the proper tooling, direction, and context.” The few times the model fell short, they add, were resolved by improving discipline and workflow, not by switching models or moving to the cloud.

The anecdote is far from trivial. It highlights a turning point for those evaluating on-premises deployment: with a quantized model (the A3B designation suggests aggressive quantization, possibly 3-bit) running locally and proving sufficient for complex technical tasks, the real gap is no longer in the model’s compute capacity but in the quality of the interaction the user is able to design. Prompt engineering, integration with external tools, and a disciplined workflow become the true multipliers of effectiveness.

This shift in the bottleneck has deep implications for the industry. If an on-premises LLM is “good enough” for many technical activities, the economic argument pushing toward cloud APIs – per-token costs, network latency, vendor lock-in – loses its bite. Total Cost of Ownership (TCO) needs recalculating: hardware purchased for local inference can be amortized over long periods, zeroing out variable costs and keeping data in-house, a clear advantage for data sovereignty and compliance. At the same time, cloud providers will have to justify added value that goes well beyond raw power, betting on integrated ecosystems or capabilities that local setups cannot yet replicate.

There is, however, a flip side. The user’s experience exposes an uncomfortable truth: many organizations may chase ever-larger models when the real leap in quality lies in refining processes. A powerful LLM used poorly breeds frustration; a modest one embedded in a well-designed workflow can be enough. The risk is that the industry keeps pouring resources into omnivorous mega-models, fueling a race that has already peaked for many use cases.

This isn’t a universal thesis – for highly specialized tasks or complex conversational needs, larger models still hold an edge – but for software development and technical planning, the signal is clear: the sufficiency threshold has been crossed. The user’s provocative question (“Is everything past this point just enabling laziness?”) strikes a nerve. We already have sufficient local tools, yet we keep chasing AI giants in the hope they’ll read our minds. Perhaps real progress lies in cultivating discipline, not the model.

For those weighing on-premises deployment, this account offers a concrete reference point. You don’t need GPU clusters costing tens of thousands of euros to get useful results: a quantized approach and solid workflow engineering can tip the balance, restoring control, cost predictability, and independence.