Open-weight models have outgrown the hardware most companies own. Teams that want to run them on their own infrastructure, for data residency, cost or latency, are left to close that gap themselves: quantization, distillation, memory management, serving strategies under a fixed hardware ceiling.
Does that work count as SR&ED? It depends on what you attempted and why, not on the technology involved.
Start with the test, not the tech
The CRA assesses the work performed. It keeps no list of qualifying technologies. The same three criteria apply as always: whether the work sought a technological advancement, whether it faced uncertainty that standard practice could not resolve, and whether it proceeded by systematic investigation.
This matters more in AI than most fields, because the tooling has become very good and much of the daily work is configuration against documented parameters. But the same activity can fall on either side of the line. Fine-tuning against a published recipe is one thing. Fine-tuning when that recipe fails against your constraint, with no way to know in advance what will work, is another.
Four questions worth asking
What constraint were you working against? Uncertainty tends to live between a requirement and available capabilities: a VRAM ceiling, a latency target, an accuracy floor that has to hold on your own data.
What did standard practice suggest, and how did it fall short? If a documented approach existed and worked, there was no uncertainty. If it measurably failed against your constraint, that failure is where the SR&ED story begins.
What did you try next, and why that? Systematic investigation means hypothesis, test, conclusion. Evals, ablations and benchmark runs are evidence of exactly this, if they connect to the question they were meant to answer.
What did you record while the work was happening? The CRA expects contemporaneous documentation and narratives describing specific uncertainty rather than general improvement. Run logs and commit history are strong evidence, and much weaker if nobody can tie them to a project six months later.
Work that did not succeed
An abandoned attempt is not a weaker claim. If a team spent six weeks trying to bring a model within a hardware envelope and could not, that outcome is direct evidence of the uncertainty the program exists to support.
Next steps
None of this determines anything about your projects; eligibility turns on your facts. Review your current AI work against these questions before your filing deadline, and check whether the evidence you already generate would hold up under review.
Recent legislative changes have also restored capital expenditure eligibility, which may affect how the hardware behind this work is treated. We cover that separately.
Leyton’s technical consultants work with Canadian AI and software teams to assess projects against the SR&ED criteria. Contact our team to discuss your current work.