Open-weight AI models from Meta and NVIDIA broaden options for local and enterprise deployment. Chinese models still lead on major benchmarks, while hardware costs, licensing and reliability remain central to model selection. Businesses gain more choice and control over how they deploy and adapt AI.
New releases give developers more options for running capable models independently, although Chinese laboratories retain important advantages. For a company choosing an AI model, a benchmark score is only part of the decision. The team also needs to know where the system can run, what its licence permits, and how much it will cost to operate. A model available through an API may be convenient, while downloadable weights offer more control over deployment and adaptation.
Announcements from Meta and NVIDIA on 15 October 2026 have added to those choices. Meta released Muse Spark Open, a large multimodal model designed for coding and agentic work. NVIDIA expanded its Nemotron family with models aimed at enterprise workflows and local deployment.
The releases strengthen the range of models available from U.S. companies. They also invite comparison with Chinese developers, whose Qwen, Kimi, GLM and DeepSeek families have become important options in the open-weight ecosystem. Independent evaluations still show several Chinese models ahead of Meta’s latest release. The immediate development is a broader set of credible alternatives, with different strengths and deployment requirements.
What Open-Weight Models Actually Provide
An open-weight release makes a model’s trained parameters available for download. Depending on the licence, developers can run it on their own infrastructure, modify it and adapt it to particular tasks. That access differs from using a closed model through a provider’s application or API. It also differs from a fully open-source release: downloadable weights do not automatically come with complete training data, training code or unrestricted usage rights.
For businesses, the appeal is practical. A manufacturer might want an assistant that searches internal engineering records. A research team may need to inspect how fine-tuning changes model behaviour. An enterprise may prefer to operate within its own network.
Self-hosting can support those goals, but it transfers responsibility to the operator. Hardware, security, software updates, monitoring and capacity planning still need to be managed. Whether it saves money depends on the workload and how effectively the infrastructure is used.
Muse Spark Open Brings a Large Model to the Desktop
Meta describes Muse Spark Open as the most capable U.S.-developed open-weight model to date. This is the company’s assessment; independent results vary by benchmark. The model has 816 billion total parameters, with approximately 42 billion active for each token. It supports multimodal inputs and a context window of up to one million tokens.
Its mixture-of-experts architecture helps explain those two parameter figures. Instead of using every parameter for every token, the model routes computation through selected experts. This reduces computation relative to activating the entire network. The storage requirement remains substantial. Apple reports that the weights would require more than 1.6 terabytes in their native BF16 precision. Its tested 4-bit version reduces the model’s weight footprint to approximately 493GB, before additional memory needed during inference.
Meta and Apple have enabled deployment across two M5 Ultra Mac Studios using Apple’s MLX framework. This is a meaningful demonstration of local inference, although it still requires a substantial hardware setup. The distinction matters: fewer active parameters can make computation more manageable without making the full model small. A published maximum context window also should not be read as a guarantee that every deployment can use that capacity at a practical speed.
NVIDIA Targets Different Deployment Needs
NVIDIA’s announcement covers several sizes and use cases. Nemotron 3.5 Super VL is a 120-billion-parameter model intended for multimodal and agentic workloads. Its capabilities include working with documents, charts and visual information, alongside tool use and multi-step tasks.
The smaller Nemotron 3.5 Bolt 74B-A9B 24GB targets local agentic applications. NVIDIA says its native 2-bit quantization allows it to fit within a 24GB memory footprint. Both models became available on 15 October. A further model, Bolt E4B 4GB, was announced for 23 October and was not yet available when this article was prepared.
These releases address different operating requirements. A local assistant and a shared enterprise service have different needs for memory, response speed and concurrent users. Buyers should assess each model and its serving configuration separately rather than treat the Nemotron family as one performance category.
Chinese Models Still Lead on Important Measures
The latest announcements do not establish that American laboratories have overtaken their Chinese counterparts. At the time of checking on 16 October, Artificial Analysis listed Muse Spark Open at 39 on its Intelligence Index. Kimi K3 scored 44 and GLM-5.3 scored 45 at the reasoning settings shown in the comparison.
Those results place Meta’s model below both on that evaluation. They also illustrate why claims about “closing the gap” need a defined benchmark, named model versions and a comparable historical baseline. An aggregate index cannot settle every deployment decision. Coding performance, document understanding, tool use, latency and resource consumption may matter differently across applications. Agent results also depend on the surrounding software, tools and inference settings.
The U.S. open-weight effort predates this month’s releases. OpenAI released gpt-oss in August 2025, while Google has continued developing its Gemma family. October’s announcements extend an existing effort rather than mark its beginning.
Licensing Must Be Checked Model by Model
Licensing differences do not follow a simple divide between American and Chinese releases. Qwen3.8-27B, for example, is listed under Apache 2.0, while DeepSeek-V4-Pro uses the MIT licence. Both offer familiar licensing frameworks for commercial development.
Muse Spark Open uses Meta’s custom licence. It includes attribution and usage requirements, along with a separate-agreement provision for certain model-as-a-service businesses whose aggregate revenue exceeds $20 million over a consecutive twelve-month period. The provision includes exemptions for specified internal and other product uses. The practical lesson is to examine the exact model and intended use. Its country of origin does not establish what a business may build, distribute or sell.
What Enterprise Teams Should Evaluate
For an enterprise, a useful comparison starts with representative work. An engineering assistant could be tested on whether it retrieves the correct revision of a maintenance procedure and cites the relevant passage. A finance assistant could be assessed on invoice extraction, reconciliation questions and its ability to recognise missing information. An agent that updates business systems needs testing around permissions, failed actions and human approval.
These checks reveal requirements that general leaderboards cannot resolve alone. Data location deserves similar care. A model developed in China can be downloaded and hosted elsewhere. A U.S.-developed model can also be accessed through an external hosting provider. Model origin and the location where business data is processed are separate considerations.
Local deployment offers control, but privacy and compliance still depend on the complete application: its retrieval system, logs, access controls, external tools and operating procedures. Cost comparisons should cover that complete system as well. API services may remain economical for occasional or unpredictable demand. Sustained workloads may justify dedicated infrastructure, provided the organisation can operate it effectively.
A Wider Choice, With Trade-offs
Meta and NVIDIA have expanded the choices available to teams that want to run and adapt capable models independently. Meta’s release demonstrates how a very large multimodal model can operate on desktop-class systems. NVIDIA offers a range aimed at different enterprise and local workloads.
Chinese laboratories remain strong competitors, including on current independent evaluations. The evidence supports a more competitive market, while leaving the question of overall leadership dependent on the task and measurement. For developers, the next step is to compare suitable models against actual work. Accuracy, operating cost, licence terms and deployment effort will determine which releases become useful infrastructure.