A Reddit post from a user who doesn't own a faster RTX has reignited a discussion that cuts to the core of local inference. The request is as simple as it is revealing: Qwen 3.8 27B works well, but in xhigh mode it takes ages on an M1 Max, to the point where a single task has to run overnight. That's why the author is asking for a Qwen 3.8 35B A3B variant, knowingly accepting a somewhat less brilliant model in exchange for answers in human timescales.

The issue is not a lack of memory, but the cost of reasoning. The user knows that the 27B's intelligence comes largely from long thinking times: this is what is known as test-time compute, the ability to use more inference steps to improve the quality of the result. On an M1 Max, however, each additional step translates into real waiting, energy consumed, and hardware locked up for hours. It's not a VRAM problem or a matter of loading weights: it's a problem of sustained throughput and model architecture.

The A3B label suggests a different path. Without official details, it plausibly indicates a model family with fewer active parameters than the total, a design typical of MoE architectures. For anyone running models locally, this means reducing computational work per token without giving up quality entirely. The user is not looking for a smarter model: he's looking for a model that doesn't force him to choose between a night of computation and the cloud.

This grass-roots request has structural implications. If long thinking becomes the main route to higher intelligence, those without fast dedicated GPUs are cut off from the practical benefits. The demand for a 35B A3B variant signals that the open-weight market will need to produce not just more capable models, but variants optimized for local inference with fewer active parameters. Alibaba and other labs have already shown they can attract the self-hosted community; now the pressure is on everyday practicality, not on a single leaderboard.

There is a second layer of consequences. From an enterprise perspective, a model that takes hours to complete a task is not compatible with interactive workflows and generates energy and waiting costs that weigh on TCO. MoE architectures are therefore not just a technical shortcut: they are a tool to maintain perceived quality without imposing a uniformly Nvidia hardware fleet. For those evaluating on-premise deployment, AI-RADAR offers analytical frameworks at /llm-onpremise for comparing these trade-offs without reducing everything to raw VRAM.

To the user's closing question — where is the future heading, toward ever longer reasoning times? — the most realistic answer is that the industry will learn to modulate the reasoning budget according to task difficulty. Simple tasks should not pay the price of complex ones. That is a change that rewards those who design runtimes and serving stacks able to adapt the length of thought, and it brings consumer hardware back into the conversation.