Comparisons
Compare architectures, deployment routes and delivery responsibilities.
- Pages
- 11
- Sources cited
- 15
- Last updated
- 26 October 2026
Which intervention deserves a test?
Keep current API
- Best initial hypothesis
- Current quality and whole-life economics are acceptable
- What changes
- Nothing material
- Buyer still owns
- Monitoring and reassessment
- Failure or stop signal
- A measured constraint emerges
These are decision hypotheses from the research. The buyer’s own task, rights and acceptance rules determine fit.
Compare the system, not the model label.
Start with a dated, reasonably optimised baseline.
Read the full research.
Which option should we test first?
Choose the option that addresses the diagnosed failure with the least irreversible work.
Prompting, retrieval, adaptation or routing?
TRL’s SFT documentation establishes a concrete implementation route for supervised adaptation.
LoRA/QLoRA versus full fine-tuning: what is the practical trade-off?
QLoRA uses quantisation with low-rank adaptation to reduce training memory.
Which delivery route fits our team?
Together AI and Fireworks describe adaptation/inference platforms; Winder.AI describes consulting and development services.
Shared API, dedicated managed deployment or self-hosting?
Modal and Baseten publicly describe different ways to serve custom workloads.
How do we normalise competing proposals?
Issue a common response schedule with rows for data curation, labels, baseline optimisation, adaptation runs, evaluation, serving integration, security review, monitoring, on-call support, retraining and transition.
What comparison could change the decision?
Run an explicit sensitivity review: reduce volume; reduce API prices hypothetically; increase human-review minutes; add required redundancy; and remove a planned discount.
Frequently asked questions
Short answers to the questions buyers ask.
Read →09 / REFERENCEWhat remains unproven
Evidence gaps and limits that still need resolving.
Read →10 / REFERENCEEditorial self-assessment
The research’s own quality scores and limitations.
Read →11 / REFERENCESources
Trace the section’s claims to the original sources.
Read →Useful evidence.
Visible limits.
Strong architecture and procurement comparison with explicit responsibilities. Performance and total-cost rankings remain deliberately unfilled without controlled buyer evidence.