Consulting · AI Systems
Legacy LLMs vs Open-Weight Models
Rent the frontier, or own the weights.
The first generation of enterprise AI ran on hosted models accessed through a vendor’s API: powerful, easy to start with, and entirely outside the firm’s control. Open-weight models, whose parameters can be downloaded and run anywhere, have closed much of the capability gap for most business tasks. The choice is no longer which is better in the abstract but which belongs where.
Hosted frontier models
The largest models from the major labs, accessed over an API and billed per token. They lead on the hardest reasoning tasks, need no infrastructure, and improve without effort on the firm’s part. The costs are structural: every prompt leaves the building, pricing and behavior change on the vendor’s schedule, the model cannot be inspected or fixed, and the firm builds on a dependency it does not control.
Open-weight models
Models released with their weights, which the firm can run on its own hardware or a private cloud, adapt to its domain, pin to a version and audit. Capability trails the very largest hosted models on the hardest tasks and matches or exceeds them on many routine ones, at a fraction of the marginal cost once infrastructure is in place. The costs are operational: hardware or reserved compute, engineers who can run inference at scale, and responsibility for safety and updates.
Where each belongs
Anything touching client data, positions, credit files or internal documents belongs on models the firm controls, either open-weight in its own environment or hosted under contractual terms that prohibit training and guarantee deletion. Hard, infrequent reasoning over public information can justify a frontier model. High-volume, well-defined tasks, classification, extraction, summarization, drafting, are usually cheaper and more controllable on an open-weight model tuned to the job.
Routing across both
Mature deployments do not choose once. A routing layer sends each request to the cheapest model that meets its quality bar, keeps sensitive requests inside the perimeter, and lets the firm swap models as the market moves. Building that layer early is what prevents lock-in later.
Practice
How Wesvel approaches it
- 01Tasks are classified by data sensitivity, volume and difficulty before any model is chosen.
- 02Hosted and open-weight candidates are benchmarked on the firm’s own evaluation set, not on published leaderboards.
- 03Total cost is modeled across tokens, infrastructure, engineering and switching, over a multi-year horizon.
- 04A routing and abstraction layer sits between applications and models from day one.
More in AI Systems
- Own Your DataKeeping proprietary data inside the firm’s perimeter, understanding what vendors may learn from it, and building retrieval so models use your knowledge without absorbing it.
- Training Your Own ModelsWhen to fine-tune or distill an open-weight model on internal data, how to do it without leaking it, and how to prove the result is better.
- Building Internal SystemsThe architecture around the model: orchestration, tools, retrieval, evaluation, monitoring, security and the infrastructure to run it.
- Governance and ComplianceModel risk management, human oversight, record-keeping, vendor diligence and disclosure for AI in regulated financial businesses.
- AI in Financial OperationsThe use cases that hold up in investment, lending and advisory operations, and the ones that do not, drawn from how Wesvel uses these systems itself.