# Local Model Adoption Is Accelerating. Govern Your Models and Accounts. > Source: > Published: 2026-11-10 00:00:00+00:00 [Back to Blog](https://certiv.ai/blog/) # Local Model Adoption Is Accelerating. Govern Your Models and Accounts. *Now Available* **Certiv already discovers and governs the cloud models your agents call. Now it does the same for models running locally on your endpoints, and flags abliterated builds. Administrators can review every model and account, mark each one sanctioned or unsanctioned, and block unsanctioned ones across the organization.** Do you know which laptops in your company are running a model locally? It’s increasingly likely that some of them are. It’s a model file downloaded from the internet and served on the device through a runtime like Ollama, LM Studio or vLLM. With the rising cost of cloud inference and more powerful endpoint hardware, companies are adopting local models on purpose: they cost nothing per token and never leave the machine. Employees are installing them too, often without anyone’s approval. In most cases that’s fine. Local models run quickly, and keeping data on the device can make data residency and privacy requirements easier to meet. But some of them have had their safety training removed, and almost none of them show up in the tools security teams rely on. A model no one can see is also a model not accounted for in an audit. That’s why we’re extending Certiv to models running locally on the endpoint, so [shadow AI](https://certiv.ai/blog/what-is-shadow-ai/) detection and runtime security cover every model your agents use, in the cloud or on the device. It starts with the two things every security team needs to govern first: the models your agents call and the accounts they sign in with. ## Local and open-weight model adoption is accelerating Open-weight models are already in production. In [Menlo Ventures’ 2026 survey](https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/) of nearly 400 U.S. enterprise AI decision-makers, open-source models handled about one in five enterprise tokens, up from 11% a year earlier, while accounting for about 4% of model spend. Developer traffic also shows this shift. On [Vercel’s AI Gateway](https://vercel.com/blog/ai-gateway-production-index-september-2026), open-weight models ran 56% of all tokens in August 2026, up from 7% in December, while accounting for 14% of spend. And inference is moving onto the endpoint. [IDC predicts](https://www.idc.com/resource-center/blog/three-forces-shaping-the-future-of-it-leaderships/) that by 2030, 50% of enterprise AI inference workloads will be processed locally on endpoints or edge nodes. ### Why adoption is accelerating - **The models are good enough.** Open-weight families like Qwen, Gemma, Llama and Mistral now handle everyday coding, writing and analysis, and new versions ship every few months. - **Laptops can run them.** Current laptops ship with neural processors and enough memory for capable models, and quantized builds shrink a model to a few gigabytes. - **Setup takes minutes.** Runtimes like Ollama and LM Studio install in one step, and most agent harnesses can point at a local model as easily as a cloud API. - **They’re cheap to run.** Open-weight models cost a fraction of what frontier models do, and run locally, there’s no per-token cost at all. - **Data stays on the device.** Prompts, code and files never leave the endpoint, and the model keeps working offline. - **They can slip past software controls.** Many organizations restrict what employees can install, but a runtime and a model file don’t need an account, an API key or a vendor review, so they can reach a laptop outside the usual approval process. ## Abliterated models remove the guardrails you assume are there An abliterated model is an open-weight model with its refusal training stripped out. Open tools do it in minutes on ordinary hardware. [Hugging Face](https://huggingface.co/models?search=abliterated) catalogs more than 9,100 models labeled abliterated as of November 9, 2026. People use them for: - Fewer refusals on sensitive requests - Unrestricted writing and roleplay - Less prompt wrangling to get an answer - Research into refusal mechanisms - Access to blocked and harmful content Most of those uses are legitimate. Security teams need to know where these models are, decide which endpoints can run them, and stop uses that violate company policy, including an insider who wants answers a frontier model would refuse. ## Local models are invisible to your existing controls **No network egress means nothing to inspect.** A local model runs on the endpoint’s own compute. Gateways, proxies and DLP never see the traffic. **No usage bill means no spend signal.** Cost monitoring tracks API consumption. A model running on a laptop doesn’t generate any. **Agents can call them with the user’s privileges.** Agent harnesses can use a local model as their backend, on the same machine, with the same access as the person running them. **A filename is not an identity.** Models are renamed, re-quantized and re-uploaded. A filename or a hub label doesn’t tell you what the model actually is. ## Certiv shadow AI detection now covers every model, cloud and local Certiv already discovers the agents, tools, MCP servers and provider accounts on each endpoint, sanctioned or not, along with every cloud model those agents call. Local models complete that picture: the same discovery and governance, now for models running on the device itself. [Scout](https://certiv.ai/scout/) inventories the models installed on each endpoint across local runtimes, including popular ones like Ollama, vLLM and LM Studio, and the agent harnesses that call them. Local models now appear on the Models page alongside cloud models, with new fields that show where each model is running, what the model actually is, whether it’s abliterated, and which endpoints are running it. ### Identify models by their contents Certiv identifies a model by the file’s contents. Each model file is checked against known public releases and labeled: - **Official:** matches a known release from the original publisher. - **Modified:** matches a known build with its refusal training removed, even when the file has been renamed. - **Unverified:** doesn’t match a known release, so it can’t be confirmed. ## Governance starts with models and accounts Before you can govern what an agent does, you need to govern what it runs on: the model it calls and the account it signs in with. Certiv now gives administrators one place to see and review every model and account in use, decide which are sanctioned and which aren’t, and block anything unsanctioned across the organization, with no policies to write. ### Sanction the models your teams can use Every model your agents have used, open-weight or proprietary, local or cloud, appears on the Models page with its lab, attributes such as open-weight or modified, where it runs, usage, endpoints, and when it was first and last seen. Review each one and mark it Sanctioned or Unsanctioned. Then set one organization-wide policy: Enforce blocks unsanctioned models on every endpoint, Audit records their use without blocking, and you choose whether the policy also covers models not yet reviewed. Every change is versioned and attributed. ### Keep work on company accounts Certiv also sees which provider account each agent signs in with, such as Anthropic, OpenAI or Cursor, so a personal account on a work laptop shows up next to the company’s. Accounts get the same review: mark each one Sanctioned or Unsanctioned, and Enforce blocks unsanctioned accounts across the organization. ### Runtime security enforces it, local or cloud Scout sits in-line between agents and the models they call, so traffic to a local model gets the same runtime security as traffic to a cloud API. Policy is evaluated before the request runs. Agent actions are governed down to individual tools, MCP servers and skills. Sensitive data and files are kept from reaching the model. ## Built for the endpoint, ready for what comes next Certiv runs on the endpoint, in-line between agents and every model they call, in the cloud or on the device. Detection is data-driven and delivered from the Certiv cloud service, so new runtimes, agent harnesses and model sources can be added as they appear. [Local model governance](https://certiv.ai/local-models/) is available now to all Certiv customers. See every model your agents use, in the cloud and on your endpoints. *- Tai, VP of Product, Certiv*