Home AIThe Future of Open Weights: What Comes Next, and What to Do About It (5/5)

The Future of Open Weights: What Comes Next, and What to Do About It (5/5)

by Vamsi Chemitiganti

This is the final part of the series. We’ve covered what open weight models are, how they compare with closed models, where each belongs in an enterprise architecture, and how a bank might put them to work. Now for the harder part: where is this heading?

My core prediction: the model is becoming a component, and the competitive advantage is shifting to the platform around it and the proprietary data flowing through it. Here is how I see that playing out — with the caveat that anyone claiming certainty about AI timelines is selling something.

How I expect the open weight landscape to evolve (my view, not a forecast of record)

Six trends I’m watching

  1. The “good enough” line keeps rising. Open weight models now trail the frontier by months rather than years on many tasks (Digital Applied). The frontier will keep moving, but the share of enterprise work that needs it is shrinking. Expect open weights to become the default for the routine 80% of workloads, with closed models reserved for the hardest problems.
  2. Smaller, specialized, everywhere. The most interesting releases this year were not the trillion-parameter giants. They were efficient Mixture-of-Experts models activating only a few billion parameters, and small dense models that fit on a single GPU or a laptop. Distillation from frontier models into compact, task-specific open models will become a standard enterprise pattern — including at the branch, the factory floor and the device.
  3. Inference economics keep improving. Serving engines such as vLLM and SGLang, plus specialized silicon, cut hosting costs dramatically over the past year. Cheaper inference shifts the bottleneck from “can we afford to run it?” to “can we govern it?”
  4. Agents make multi-model the norm. An agentic workflow may call a dozen models in a single task: a frontier model to plan, small open models to execute, classifiers to check. That makes the gateway and routing layer from Part 3 more critical, not less — and it makes per-call cost and latency of the worker models a first-order design concern.
  5. Provenance becomes a board-level question. Many of today’s leading open weight models — Qwen, DeepSeek, Kimi, GLM — come from Chinese labs. For regulated and public-sector buyers, where a model was trained, on what, and under which jurisdiction is now a procurement and risk question, not just a technical one. I explored the broader version of this in my recent piece on AI sovereignty. Expect model bills of materials (AI-BOMs), signed checkpoints and provenance attestations to move from nice-to-have to required.
  6. Regulation catches up — unevenly. The EU AI Act already places obligations on general-purpose model providers, with only partial relief for open releases. US banking regulators have explicitly deferred generative AI from their revised model risk guidance while they gather input (Orrick). Firms that build disciplined governance now will be ahead of whatever comes next, rather than retrofitting it.

What could go wrong

It’s worth stating the risks plainly. Open weight releases depend on the strategic choices of a small number of labs; a major publisher could tighten its license or stop releasing weights. Supply-chain attacks on model hubs are a real and growing threat. And many enterprises will underestimate the operational cost of running models themselves. None of these are reasons to avoid open weights — they are reasons to architect for optionality.

An 18-month agenda for the CIO

  1. Build the gateway first. One governed entry point for every model call, with routing driven by data classification and task type.
  2. Stand up a model registry and supply-chain controls. Scan, sign and pin every checkpoint; treat weights as production artifacts.
  3. Pick two or three high-volume, document-heavy workloads and move them to fine-tuned open weight models inside your boundary.
  4. Invest in evaluation as a core competency. Domain-expert golden datasets are what let you swap models on evidence rather than vendor claims.
  5. Pool GPU capacity across use cases and track utilization the way you track database performance.
  6. Keep a frontier relationship. You will still need the best closed models for the hardest problems — just not for everything.
  7. Put provenance and licensing in procurement. Legal and risk should sign off on every model’s license and origin, not just its benchmark scores.

Closing thought

When I started this series, I said the open vs. closed debate is a false binary. Five posts later, I’ll put it more directly: the winners won’t be the enterprises that pick the right model. They will be the ones that build the platform that makes the choice of model reversible. Open weights are what make that reversibility real.

Thanks for reading the series. As always, I’d welcome your comments and pushback.

Sources

Featured image designed by Magnific

Disclaimer

This blog post and the opinions expressed herein are solely my own and do not reflect the views or positions of my employer. All analysis and commentary are based on publicly available information and my personal insights.

Discover more at Industry Talks Tech: your one-stop shop for upskilling in different industry segments!

Ready to master the future of telecom? My book, “Cloud Native 5G – A Modern Architecture Guide: From Concept to Cloud: Transforming Telecom Infrastructure (Industry Talks Tech)” is now available on Amazon.

You may also like

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.