Companies that want to use AI agents for their own processes have so far mostly faced the same decision: either use a cloud provider's interface and let data leave the house, or forgo current model quality.Two publications from 15 October 2026 shift this trade-off.Nvidia is expanding its Nemotron family with an open agent model, and Meta is back with an open-weight model for local operation.

Nemotron 3.5 Super VL: agents with an eye for documents

With Nemotron 3.5 Super VL, Nvidia is releasing an open multimodal model with 120 billion parameters.Multimodal here means: the model processes not only text but also images.The stated purpose is enterprise agents – that is, software that works through tasks independently in multiple steps instead of merely answering individual questions.

The combination of text and image in particular is interesting for business processes.Scanned delivery notes, screenshots from specialist applications, forms with a table structure or product photos can in principle be captured in the same processing step as the associated running text.

In addition, Nvidia is providing two compact models under the name Bolt.They are intended to bring agents to local machines – that is, to workstations or servers in your own network, instead of rented computing time from a provider.That the concept resonates is shown by its reuse: according to the report, dozens of organisations are already fine-tuning Super VL, adapting it with their own data to their specialist domain.

However, the report explicitly names one limitation: anyone who wants to use the model optimally ends up with Nvidia hardware.In this case, open weights therefore mean independence from the provider at the software level, not necessarily at the infrastructure level.

Muse Spark Open: large context windows, Apple as target platform

Meta is taking a different approach with Muse Spark Open.The open-weight model – that is, a model whose weights are freely available – comprises 816 billion parameters and processes text, image, video and audio.Context windows of up to one million tokens are supported.The context window describes the amount of material the model can keep in view at the same time; one million tokens corresponds to a very extensive collection of documents in a single run.

According to the report, Muse Spark Open is roughly on a par with the most powerful open Chinese AI models in benchmarks.It is optimised for local operation on Apple's M5 Ultra.

Here too, the same limitation applies as with Nvidia, only with the opposite sign: the model is open, but the performance advantage is tied to a specific hardware platform.

What the two announcements show in common

Both publications point in the same direction but set different priorities:

  • Scale: 120 billion parameters for Super VL versus 816 billion for Muse Spark Open.The smaller model is therefore much closer to hardware that is realistically available in companies.
  • Modalities: Super VL covers text and image, Muse Spark Open additionally video and audio.
  • Target picture: Nvidia explicitly addresses enterprise agents and, with the Bolt variants, supplies compact models for local machines.Meta emphasises model breadth and context length.
  • Hardware dependency: Nvidia accelerators on one side, Apple's M5 Ultra on the other.

The commonality is more decisive than the differences: powerful models are no longer available exclusively as a cloud service.For data protection, ongoing costs and dependence on a single provider, this changes the starting position.

Assessment for practice

No conclusions about operating costs, licence details or concrete response times can be drawn from the information available.Anyone preparing a decision should therefore look less at parameter counts than at three points.

First, the procurement question. Local operation always also means: hardware, electricity, operation and maintenance remain in-house.The hardware dependency of both models makes this an investment decision, not just a software question.

Second, the fit to the use case. For processing documents with text and image components, a more compact model geared towards agents is often the more practicable choice than the largest possible model.The fact that Super VL is already being fine-tuned by numerous organisations speaks for a growing base of practical experience.

Third, the architecture of your own application. Anyone who encapsulates model calls behind a clearly defined internal interface can switch models later – between cloud and local operation as well as between providers.This decoupling is the most effective safeguard against a commitment you regret in two years.

For web and software projects, this means: the question in future is less often "cloud or not at all" and more "which model at which point".Interfaces, data flows and permissions should therefore be designed from the outset so that a model change is not a rebuild of the application.

Anyone planning an integration today gains more from clean abstraction than from committing to a specific model.

Sources