# Introducing Cisco Model Foundry: Bringing AI to our partners > Source: > Published: 2026-10-26 19:59:44+00:00 A year ago, the answer to “which model should we use?” was simple: whichever frontier model topped the leaderboard that week. Then open models caught up. They cost a fraction as much, you could run them where you wanted, and your data stayed yours. Plenty of teams made the switch. Then they pointed those models at real work. An operator reports that R2 lost every route from R1, but ping across the link still works. A security team needs to find the three files in a million-line codebase most likely to hold a vulnerability. An observability team needs to spot the one interface that matters in a week of telemetry. This is where general-purpose models, open or closed, start to struggle. We think the next leap comes from a model that already knows the domain and then learns your customer. That second step runs on your data and your relationships, and it’s where partners lead. Today at Partner Summit, we’re introducing **Cisco Model Foundry**, a new platform service in Cisco Cloud Control. It gives partners access to Cisco’s purpose-built models and lets them serve those models in customer applications. When the use case calls for it, partners can also fine-tune the models on their customers’ own data, all in one place.  ## Access the expertise. Put it to work. Tailor it when it counts. **Access Cisco’s domain models.** Model Foundry opens up the models Cisco has built on four decades of networking, security, and operations knowledge. Several of them previously ran only inside our own products. Browse the catalog, open the Experiment, paste in a real problem, and see how each model handles it before you commit.  **Put them to work.** In Model Foundry, you create an API key, set an expiration, choose a model, and you’re connected. The API is OpenAI-compatible, so it plugs into the agents, tools, and workflows you already run. Administrators get tenant-scoped keys, per-model access controls, and full visibility into usage by model, key, and day.  **Tailor it when it counts.** Fine-tuning means taking the knowledge that makes a network or a business unique and training the model on it. It’s built for the large volumes of data that partners and customers already have: configs, log messages, alerts, tickets, and agent traces from thousands of devices. Model Foundry handles the hard parts, from the GPUs to the training pipeline to built-in evaluation that runs the base and tuned models side by side. You see where tuning helped, and where it didn’t, before anything reaches production.  Many partners don’t need to fine-tune to get value. A purpose-built Cisco model, served as-is, is built to outperform generic open models on domain work. Fine-tuning is there for the moments when “good” isn’t good enough. ## Models that already speak the language of infrastructure Model Foundry starts with models that already understand your world: - **Deep Network 1.4.** Built for network operations: investigation, root-cause analysis, and multi-vendor troubleshooting. It was trained on networking telemetry, CLI syntax, configuration data, and troubleshooting topologies, the kind of knowledge a CCIE relies on, in a fraction of a frontier model’s size. - **Antares-3B.** The most capable model in Cisco’s Antares family, purpose-built for agentic vulnerability localization. It pinpoints the source files most likely to contain a specific vulnerability. - **Decision Model 1.0.** A fast decision layer for AI agents working on networks. It picks the right next tool to call, checks whether the evidence supports a diagnosis, and knows when to escalate to a human. - **Cisco Time Series Model 1.0.** Built for the operational and observability data that infrastructure generates every second. These models sit alongside a growing library that includes the Foundation-Sec family, SecureBERT 2.0, and Cisco’s open-weight Antares models, so partners have a real catalog to evaluate against their customers’ tasks. ## The proof: WWT We tested Model Foundry with World Wide Technology (WWT). WWT evaluates many agent harnesses, including agentic troubleshooting harnesses. An operator describes a problem. The agent investigates the live network with tools, finds the root cause, cites its evidence, and recommends a fix for a human to approve. It’s a demanding job, and the Deep Network Model, while strong, had never seen WWT’s network. Cisco fine-tuned the 31-billion-parameter Deep Network Model on agent and troubleshooting data and evaluated models in the lab, on real Cisco devices, using the same agent, tools, and settings, against problems none of the models had seen. The problems spanned BGP, OSPF, link failures, router policy, simultaneous faults, and harmless config changes designed to look like faults. Here is how often each model recommended the exact right fix: ##### The difference shows up most clearly on a single alert. The syslog points in the wrong direction, and the base model follows it. The fine-tuned model sets the misleading clue aside, pinpoints an OSPF dead-timer mismatch on the right interface and VLAN, and tells the operator how to fix it. That’s the kind of judgement fine-tuning can add. We did this in five weeks, and there’s more to do. Recognizing a healthy network when nothing is broken and untangling two faults at once are still hard problems. They’re our next targets, and each round of production data gives us more to learn from. *“By utilizing a model that understands specific environments and does more than just general networking, we were able to see real results, fast.” said Naasief Edross, Chief Security Architect at World Wide Technology. “Fine-tuning Cisco’s Deep Network Model on troubleshooting data demonstrated the ability to get to the right fix in 11 of 12 real-world scenarios in the lab. The Cisco Deep Network Model was trained to become better at identifying misleading clues and finding the real root cause of an issue. That’s what we’re after as we bring agentic operations to our customers.”*  ## Why partners, and why first The market is moving from “buy the best model” to “apply the right one.” That second question is where partners create value Cisco can’t create alone. We build the models. Partners know the customer’s environment, workflows, and history, and they hold the data that makes a model truly theirs. That’s why Model Foundry launches with partners first. Partners can build a business on it in three ways: - **Serve Cisco models as-is.** Embed purpose-built domain intelligence directly into customer applications. - **Customize where it matters.** Fine-tune a model for one customer’s network, security dataset, or workflow, and ship it. - **Build a practice around it.** Running inference is easy: get your API key and go. From there, turn fast deployment into an ongoing service, packaged with the AI infrastructure you already deliver. The result is a model built on the partner’s relationship with the customer, not just on ours. It brings together Cisco’s expertise and the partner’s knowledge, and that combination is unique to each partner and customer. ## Built into the harness Model Foundry lives inside [Cisco Cloud Control, the secure harness for the agentic era](https://blogs.cisco.com/ai/cisco-cloud-control-the-secure-harness-for-the-agentic-era). Every model you serve or fine-tune in Model Foundry gets its own Open-AI compatible API endpoint, so it can power any agent, application, or workflow you and your customers already run. Access is managed through tenant-scoped API keys, with usage visible by model, key, and day. Your training data and the models you create belong to you.  ## What comes next For decades, Cisco has helped build the networks the world runs on. Now we want to help build the intelligence that runs on them, with partners who know our customers best. Our hope is that partners rely less on rented intelligence and more on models they own, shaped to each customer they serve. While Model Foundry is our first-party home for Cisco’s specialized models, and for our open models, we’re also expanding access through an initial early-access cohort of partners: Armada, AWS, Baseten, Distyl, Fireworks, Microsoft, Ollama, RiverAI, and WWT. Through hosted inference, modular data centers, developer tools, as well as hands-on customization and solution design support, developers and enterprises have more choice in where and how they use Cisco’s open models. Participating partners receive early access to weights so they can prepare for each new release. We also plan to bring select proprietary Cisco models to these ecosystem partners in the coming days. Cisco Model Foundry enters Controlled Availability at Partner Summit, with partners getting access first. Come see it live in the keynote and at the booth. We’re glad you’re here for it. *Some products or features described may be in various stages of development and offered on a when-and-if-available basis.*