OpenHands 將 NVIDIA Nemotron 3.5 Super VL 引入軟件工程智能體工作流

TL;DR: As an NVIDIA launch partner, OpenHands is bringing Nemotron 3.5 Super VL, a new open multimodal model, into always-on software engineering workflows.

AI coding agents are moving from one-off prompts to long-running systems. The first wave proved that agents can write code, inspect repositories, run commands, fix bugs, and help developers move faster. The next wave is broader: agents that run in the background, respond to events, complete multi-step workflows, and operate across teams, repositories, documents, tools, and environments.

That shift changes the model requirements. It is no longer enough for an agent to generate code from a text prompt. Agents need to gather context, inspect artifacts, use tools, make decisions, and continue working across multi-step workflows. Increasingly, that context is not just code or text. This is where multimodal models become important for software engineering agents.

NVIDIA Nemotron 3.5 Super VL is an open multimodal model for long-running agents that need to see, reason, and act. It combines text, image, document, and video understanding with agentic reasoning and tool use in one model.

As an Nvidia launch partner, OpenHands is bringing Nemotron 3.5 Super VL into the agent infrastructure developers and enterprises need. The future of agent infrastructure isn’t one model, one prompt, or one interface. It’s systems of models connected to the tools, repositories, runtimes, policies, and workflows where engineering work happens.

Multimodal models make agents more capable. OpenHands makes those capabilities operational.

Always-on agents need more than text

A coding agent does not just answer a prompt. It reads files, searches a repository, edits code, runs tests, interprets errors, calls tools, and decides what to do next. Sometimes it opens a pull request. Sometimes it responds to a GitHub issue, Slack message, CI failure, or scheduled automation. Sometimes it needs to continue working after the developer has closed their laptop.

But many engineering workflows break out of the text-only frame. A bug report may include a screenshot, logs, and reproduction steps. A support escalation may include a customer document or screen recording. An incident may require interpreting charts, dashboards, and written updates. A UI review may require comparing implementation against a design. A migration may involve diagrams, PDFs, tables, and internal documentation.

If agents are going to handle more of the software development lifecycle, they need to understand these artifacts natively. That is the role Nemotron 3.5 Super VL is designed to play: a high-accuracy multimodal execution inside a broader system of models.

Nemotron 3.5 Super VL brings visual understanding into agent execution

NVIDIA Nemotron 3.5 Super VL is built on a 120B-parameter mixture-of-experts architecture with 12B active parameters. It supports text, images, documents, and video as input, with text output.

The important part is not just that the model is multimodal. It is that the multimodal capability is paired with agentic reasoning and tool use.

That combination matters because agents need to move from understanding to action. A model that can read a document, interpret a UI, or understand a screenshot is useful. A model that can do that while preserving context across a multi-step workflow and using tools to complete the task is much more useful.

NVIDIA positions Nemontron 3.5 Super VL as the high-accuracy multimodal execution model in a system of models, alongside Nemotron 3.5 Lightning for high-volume text execution and Nemotron 3 Ultra for the hardest text-only reasoning.

That system-of-models framing is important. The future of agent infrastructure is not one model for every task. Some tasks need fast text execution. Some need frontier reasoning. Some need native visual and document understanding. The agent platform has to route work to the model that fits the task, policy, environment, and cost profile.

That is exactly how OpenHands is built.

OpenHands gives teams the platform to run multimodal agents their way

OpenHands is the open-source platform for building and running software engineering agents. More specifically, OpenHands is where model capability becomes an agent system.

A model by itself does not review a PR, fix a CI failure, triage an issue, or update docs on a schedule. It needs a harness, tools, runtime, sandbox, repository access, integrations, policies, auditability, and a way to keep working beyond a single local session. That is the layer OpenHands provides.

With OpenHands, developers can start locally in Agent Canvas, run agents against real repositories, connect tools, and experiment with workflows. As those workflows become useful, teams can turn them into automations that run on schedules or respond to events from systems like GitHub and Slack. At enterprise scale, organizations need the control layer around those workflows: access control, observability, audit logs, cost visibility, and deployment flexibility.

Nemotron 3.5 Super VL fits naturally into this architecture because it gives OpenHands users another model option for workflows where agents need to understand more than code or text.

For developers, that means more context-aware agents. Visual and document understanding expands the kinds of work an agent can take on: interpreting screenshots, reading complex documents, reasoning across GUIs, reviewing diagrams, and understanding video or visual evidence.

For platform teams, it means more flexibility in the model layer. OpenHands is model-agnostic by design, so teams can choose the models that fit the task, cost profile, accuracy requirement, deployment environment, and policy constraint.

For enterprises, it means more control over specialized agent workflows. Nemotron 3.5 Super VL can be customized with NVIDIA NeMo using an organization’s data, tools, policies, and workflows. OpenHands provides the agent platform and control layer to run those workflows across teams, repositories, and environments.

The journey from local experimentation to production agent infrastructure should not require switching tools, rewriting workflows, or giving up control over where models and agents run.

As an NVIDIA launch partner, OpenHands is helping bring Nemotron 3.5 Super VL into the engineering workflows where multimodal understanding can make agents more useful: issue triage, UI review, incident analysis, documentation, support evidence, and long-running automations.

Technical foundation for enterprise agent workflows

Nemotron 3.5 Super VL is designed for organizations that need both capability and control.

Organizations can post-train Nemotron 3.5 Super VL with NVIDIA NeMo using their own data, tools, policies, and workflows. That matters for enterprise agents because generic behavior is rarely enough. Agents often need to understand internal terminology, domain-specific documents, policy constraints, tool outputs, and established engineering practices.

NVIDIA also highlights production deployment with NVIDIA NIM, which packages the model as an optimized microservice across cloud and data center environments.

The model is also optimized for long-running, context-heavy workloads. NVIDIA cites C-RADIOv4 vision encoding, a hybrid Mamba-Transformer architecture, multi-token prediction, and NVFP4 KV cache quantization. In practical terms, the goal is to support efficient output generation and more concurrent long-context sessions by reducing KV cache pressure and recomputation in multi-turn workflows.

For agent builders, these details matter. Long-running agents are not single-turn demos. They accumulate context, inspect multiple artifacts, call tools, and operate across many steps. Efficiency, concurrency, and context retention directly affect how practical those workflows are in production.

What Nemotron 3.5 Super VL brings to OpenHands

Nemotron 3.5 Super VL is especially relevant for workflows where agents need to interpret visual or document-based context before acting.

In OpenHands, that can look like:

• Investigating a bug report with screenshots, logs, and reproduction steps.

• Reviewing a UI change against a design or product spec.

• Extracting requirements from PDFs, tables, charts, and architecture diagrams.

• Summarizing incident evidence from dashboards, documents, and updates.

• Interpreting visual context for QA, support, or incident review.

• Researching across websites and validating information from multiple sources.

• Turning repeatable review, triage, and documentation work into automations.

These are not just chat interactions. They are workflows. OpenHands gives teams the orchestration layer to run them without locking the entire agent system to one model or one deployment pattern.

The agent stack is becoming model-agnostic infrastructure

The market is moving past the idea that AI coding agents are single-user productivity tools. Agents are becoming infrastructure. They need to run across teams, repositories, triggers, tools, and environments. They need to be observable. They need to be governed. They need to support different models for different classes of work.

That is OpenHands’ role in the market. OpenHands is not trying to be another closed coding assistant. It is the open-source agent platform for teams that want to build on an inspectable foundation, bring their own models, control their execution environments, and move from local experimentation to production agent workflows.

Nemotron 3.5 Super VL strengthens that foundation by adding a powerful open multimodal model option for agents that need to see, reason, and act.

OpenHands is proud to support NVIDIA Nemotron 3.5 Super VL and help bring multimodal model capability into the software engineering workflows where agents are becoming real systems.

About OpenHands

OpenHands is the open-source platform for building and running AI coding agents, with the interface, automations, and control layer needed to go from a single local agent to a system running across an entire organization.

Developers can start locally with Agent Canvas, connect the models and tools they already use, and run agents against real repositories. As workflows mature, teams can turn one-off agent wins into repeatable automations that run on schedules, respond to GitHub and Slack events, review PRs, fix CI failures, update docs, and keep working when a laptop is closed.

For organizations scaling agent usage, OpenHands Enterprise adds the control layer required to run agents safely across teams, repositories, and environments.

Download Agent Canvas to start running agents locally, automate real engineering workflows, and scale them when you’re ready.