By the end of this guide you will have a working definition of an open-weight model, understand why they moved from developer experiments to enterprise production in 2026, and know the five questions that decide whether your organisation should use one, and how.
The timing is not accidental. On 19 October 2026, Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model built for enterprise agents, and named EY, CrowdStrike, Salesforce and Ramp among its users. For a COO or IT Director, the question is no longer whether open models are good enough. It is which workloads belong on them.
What is an open-weight model?
An open-weight model is an AI model whose trained parameters, the weights, are published so that anyone permitted by its licence can download, run and adapt it on their own infrastructure or through a provider of their choice. Closed models such as the flagship GPT and Claude models can only be used through their maker's service.
Open-weight is not the same as open-source. FullStack’s explainer notes that an open-weight release provides the parameters and sometimes inference code, while open-source AI in the stricter sense also provides the materials and rights needed to study and rebuild the system. Computerworld adds that open-weight models are the most common kind of “open” model in enterprises today.
--- Closed model: you rent capability through an API; the vendor controls version, pricing and data handling.
--- Open-weight model: you choose where it runs, which version you keep and how you tune it, within the licence.
--- Open-source model: open weights plus training transparency, which few enterprise-grade models fully offer.
Why are open-weight models gaining ground in enterprises in 2026?
Open-weight models are gaining ground because they now handle a large share of routine enterprise work at a fraction of the price. Gateway data published in August 2026 showed open models carrying close to a third of token volume while accounting for under 4% of spend, and agents multiply that cost gap.
Constellation Research, analysing Vercel’s AI Gateway data, reported that open-weight models rose from about 11% of token volume in April 2026 to 29% in June, when roughly one in eight enterprise gateway customers ran at least one open-weight model in production. On 22 August the share hit a record 62%.
The economics explain why. In June, open-weight models produced 29% of tokens but less than 4% of spend, while Anthropic produced 32% of tokens and 61% of spend. Constellation also found that teams sending more than 10 million requests a month used an average of 35 models.
Agents sharpen the case. A single business process may involve 20 or 50 model calls, so a lower per-call price compounds. This is the same dynamic behind the inference paradox, where cheaper tokens still produce bigger bills.
How does an enterprise actually use an open-weight model?
Enterprises use open-weight models in three ways: through a managed inference provider that hosts the model for them, inside their own cloud account, or on their own hardware. Most start with managed inference, because it delivers model choice and lower cost without the burden of running GPUs, patching and scaling.
Option 1: managed inference
A provider hosts the open model and you call it through an API, much as you would a closed model. Constellation’s view is that enterprises will adopt open weights faster than they adopt self-hosted inference, and this route is why.
Option 2: your own cloud account
You deploy the model on cloud GPUs you control, which keeps prompts and outputs inside your environment and lets you fix the version. You now own uptime and security.
Option 3: on-premises or on-device
Nvidia’s smaller Bolt models are aimed at AI PCs and workstations. This suits highly sensitive work, but only organisations with real infrastructure skills should treat it as a default.
When should an enterprise choose an open-weight model over a closed one?
Choose an open-weight model when the workload is high-volume and well-defined, when data control or version stability matters more than peak reasoning, or when you need to tune the model on your own data. Keep closed frontier models for complex reasoning, ambiguous tasks and situations where an error is expensive.
Five questions settle most decisions:
--- Workload: is it repetitive classification, extraction or summarisation, or open-ended judgement?
--- Data sensitivity: must prompts and outputs stay within your own environment or region?
--- Stability: do you need the exact same model behaviour for audit purposes over 12 to 24 months?
--- Customisation: will tuning on your documents materially improve results?
--- Capability: do you have people who can evaluate, secure and monitor a model you operate?
Three or more “yes” answers point towards an open-weight option, usually through managed inference first. One or none points towards staying with a closed API for now.
Two answers deserve extra weight. If data sensitivity is high, the location and contract terms of whoever runs the model matter as much as the model itself. If operating capability is low, a managed provider is not a compromise; it is the responsible choice.
What does an open-weight deployment look like in practice?
In practice, most enterprise deployments start small: one high-volume, well-understood workflow moves from a closed API to an open-weight model through managed inference, with a closed model kept as fallback. Cost and quality are compared side by side for a quarter before any wider switch is approved.
A Hong Kong insurer's claims triage
Consider a mid-sized insurer receiving 6,000 claim emails a month in English, Traditional Chinese and mixed language. The first step, classifying each email and extracting policy number, claim type and urgency, is repetitive and well defined. That step moves to an open-weight model hosted by a managed provider in an Asian region.
The harder step, drafting a response to a disputed claim, stays with a frontier closed model. Every output from both tiers is logged against the same evaluation set of 300 historical claims, so the claims director can see accuracy and cost per claim on one page.
A logistics group's internal knowledge search
A logistics group with 40 years of operating manuals wants staff to ask questions in plain language. Because the manuals include customer contract terms, the team runs a tuned open-weight model inside its own cloud account, fixes the version for 18 months, and reviews answers monthly with operations supervisors.
In both cases the decision rests on workload shape and data sensitivity, not on ideology about open versus closed.
What does this mean for Hong Kong enterprises?
For Hong Kong enterprises, open-weight models offer a route to capable AI that does not depend on a single overseas vendor's service availability, pricing or regional policy. They also make it easier to keep personal data under tighter control for Personal Data (Privacy) Ordinance purposes, provided the deployment is properly governed.
Several US frontier providers do not list Hong Kong as a directly supported region, so many local firms reach those models through cloud platforms or group contracts. An open-weight model running in a cloud region or data centre of your choice removes one layer of that dependency.
Model origin is now part of the decision. Constellation notes that Chinese labs, including the Qwen, DeepSeek and GLM families, set much of the pace in open weights. Boards should weigh licence terms, provenance and supplier policy alongside benchmark scores. UD’s analysis of AI model concentration risk explains why a multi-model posture is itself a resilience measure.
How do open-weight models fit into a multi-model strategy?
Open-weight models fit best as one tier in a multi-model architecture, where a routing layer sends each task to the cheapest model that can do it well. Frontier models handle the hardest reasoning, open models handle repetitive agent steps, and small specialised models handle classification or extraction.
--- Tier 1, frontier closed models: complex analysis, negotiation drafts, ambiguous customer cases.
--- Tier 2, open-weight general models: summarisation, internal search answers, routine agent steps.
--- Tier 3, small tuned models: document classification, data extraction, language detection.
The control point is the router and its evaluation set. UD’s guide to LLM model routing covers how to design that layer so cost savings do not come at the expense of quality.
A practical rule for budgeting: assume the open-weight tier will carry the majority of tokens but a minority of spend, as the Vercel data suggests, and hold the frontier tier to the tasks where its accuracy measurably changes the business outcome. Review the split every quarter as model prices and capabilities move.
What are the most common mistakes with open-weight models?
The most common mistakes are treating open-weight as open-source, ignoring licence terms, assuming self-hosting is automatically cheaper, skipping evaluation on your own data, and downloading model files without verifying their source. Each turns a cost or control advantage into a new risk.
Ignoring the licence. Some open-weight licences restrict commercial use above a size threshold or for certain purposes. Legal review belongs before deployment, not after.
Assuming self-hosting is cheaper. GPUs that sit idle 70% of the time can cost more than a managed API. Compare cost per completed task, not cost per token.
Skipping evaluation. Public benchmarks do not reflect your contracts, your languages or your customers. Test on 200 to 500 real examples, including Traditional Chinese and mixed-language content.
Unverified downloads. Model files are software. Pull them only from the publisher’s official repository and record the exact version you run.
Open-weight models: key facts
--- Definition: a model whose trained weights are published for download and adaptation under a licence.
--- Latest enterprise launch: Nvidia Nemotron 3.5 Super VL, 120 billion parameters, released 19 October 2026.
--- Adoption signal: open weights rose from about 11% to 29% of Vercel AI Gateway token volume between April and June 2026 (Constellation Research).
--- Cost signal: 29% of tokens but under 4% of spend in June 2026.
--- Typical entry route: managed inference first, self-hosting only with proven operating capability.
What is the strategic takeaway?
Open-weight models are no longer a developer curiosity. They are becoming the economic workhorse for high-volume enterprise AI, while frontier closed models remain the specialists. The winning posture is not open versus closed. It is knowing which work belongs where, and keeping the freedom to move it.
For Hong Kong leaders, that freedom is also a resilience and data-control decision. We understand AI. We understand you. With UD by your side, AI never feels cold.
Now that you have the framework, the next step is mapping your own workloads to the right model tier. We'll walk you through every step, from workload assessment and model selection to secure deployment in Hong Kong and ongoing performance tracking.
Reviewed by the UD enterprise AI team.