Meta Reaches the Frontier With Open and Closed AI Models

Meta’s open-weight turnaround

This week marked a significant shift in the position of American open-weight models relative to the frontier. Meta led the change with Muse Spark Open, which became the strongest American-made open model by a substantial margin. Reflection also introduced its first model, Beam.

Without these releases, more than 20 open-weight models developed and released by Chinese companies would have ranked above the strongest Western models. Mistral has also announced Mistral Large 4, which has an index score of 38, although its weights had not yet been released at the time of writing. Additional American open-weight models are also rumored to be in development.

The performance gap between the strongest American and Chinese models appears to have narrowed by roughly half. American open models have moved from being largely irrelevant to consumers seeking frontier performance to being somewhat behind, but clearly relevant.

The development also illustrates the intense competition surrounding frontier models. Reflection and Mistral announced model scores, while Meta released the weights for a more capable model before the other laboratories could publish their own weights. All three releases are competing with Chinese models that appeared months earlier, including GLM-5.3 and Kimi K3, released on August 18 and July 16, respectively.

This article focuses on Meta’s turnaround, the significance of Muse Spark Open for the open-model ecosystem, and the balance of power among Western frontier laboratories.

Why Muse Spark Open could gain adoption

Analysis of open models has repeatedly shown that Western open-weight models receive considerably more adoption than Chinese models with a similar performance-to-size profile, even when the Chinese models are released first. Muse Spark Open arrives as open-ended large language model agents are becoming practical, following the capabilities demonstrated by GLM 5.2 and later systems.

The open ecosystem has also benefited from the rapid improvement of Chinese models. Inference providers are expanding, an open post-training economy is forming, and many enterprises want to control their own AI capabilities. Muse Spark Open performs above GLM 5.2, near DeepSeek V4.1 Flash, and below GLM 5.3 and Kimi K3. That places it within the range of models that could be considered broadly capable enough for these uses.

According to the model card published on Hugging Face:

Muse Spark Open is an 816B total/42B active parameter open weights model with a context window of 1M tokens. Built for coding and agentic work, Muse Spark Open is natively multimodal. This open weights version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails. We recommend deploying Muse Spark Open as part of a broader AI system, with additional guardrails suited to the specific use case.

The model is positioned for coding and agentic applications, supports multimodal input, and has a context window of one million tokens. It contains 816 billion total parameters, with 42 billion active parameters.

The model’s additional safety training led some observers to speculate that its capabilities had been damaged during the process, particularly because it was released after Muse Spark 1.3. A more likely explanation is that the release is a safety-tuned version of Muse Spark 1.2, which would represent only a one-point decline on the Artificial Analysis Intelligence Index while improving safety behavior.

When Meta released Muse Glimmer, a 30B agentic model, Alex Wang said the company was preparing to release Muse Spark 1.2 as open weights. That statement is consistent with the explanation that Muse Spark Open is based on that model.

The constraints of Meta’s license

The main factor that could limit adoption is Muse Spark Open’s license. It includes several restrictive provisions that have appeared in other open-model licenses:

1. Products using the model must display the phrase “Built with Muse.” Meta used a similar requirement with Llama, and Kimi has attempted one as well, although compliance has generally been limited.
2. Companies generating more than $20 million per month across their entire business while deploying the model must obtain a separate license from Meta.
3. Users must agree to an acceptable-use policy. Under stricter definitions of open-source software, this means the model does not qualify as open source.
4. These provisions also apply to model outputs and derivatives, creating legal uncertainty around uses such as synthetic-data generation.

The license does permit internal research and applications.

The practical effect of the license remains uncertain. It is complex, has not been tested in court, and many developers may choose to ignore parts of it. Numerous users are still likely to build on Muse Spark Open through different inference providers, while Meta may focus enforcement on particularly serious violations.

Licensing has not generally determined AI adoption, but that could change. Complex, nonstandard licenses may therefore create a strategic risk for laboratories trying to expand adoption of open models.

Meta’s position among frontier laboratories

A few days before releasing Muse Spark Open, Meta introduced its closed frontier models, Muse Wonder and Muse Wonder Full. On conventional benchmarks, these models are highly capable. They rank behind Anthropic’s Opus 5.5 and close to OpenAI’s GPT-6-Astra and Google’s Gemini-4-Argon.

These releases arrived slightly more than a year after Meta’s major AI redirection, which was associated with CEO Mark Zuckerberg’s Personal Superintelligence post. Meta is currently the only American company with frontier-level open and closed models. Alibaba’s Qwen has pursued a similar strategy, but with a more open orientation.

The initial evaluations of Meta’s recent models have been strong. However, expectations were that they would not immediately match the refinement of the newest Anthropic and OpenAI models in everyday use.

Meta’s recent performance reflects a distinct approach. The company is developing its own products and directions around the Muse agent while also competing for leadership in American open-weight models. Whether that strategy proves financially successful, it represents a continuation of Meta’s historically distinctive position in the industry.

The group of laboratories challenging Anthropic and OpenAI has changed repeatedly. Gemini and xAI remain potential competitors, although their offerings currently appear less unified than Meta’s. Zuckerberg and Alexandr Wang have described a model centered on high ownership and substantial computing resources per researcher. The organization is relatively flat, with Alex Wang reportedly having more than 200 direct reports, and it has produced consistent model and product improvements.

Building frontier models requires resources, computing capacity, data, and talent, all of which depend on substantial funding. Organizing those resources is more difficult. Management layers can slow decision-making and interfere with model development. Meta’s earlier Llama problems have sometimes been attributed to excessive management and model decisions influenced by internal politics rather than performance. xAI faces a different risk, involving talent turnover or insufficient organization to sustain strong models over time.

Over the past several years, Meta, Gemini, and xAI have each produced models near the performance frontier. OpenAI and Anthropic have been distinguished by greater consistency in maintaining that position. Anthropic appears particularly organized around a clear plan and vision, while OpenAI has been less consistent but has remained competitive in deployed models and has led several public scientific advances, from Navier-Stokes work to the inference-time scaling demonstrated by o1.

The current ranking described in the analysis is:

Anthropic > OpenAI > Meta approximately Gemini > xAI approximately Chinese open weights > U.S. open weights

Xiaomi V2.6 Pro represents the leading open-weight model overall in this comparison.

Open questions for the AI industry

The frontier-model business may be entering a period in which only a small number of organizations can continue competing. The industry remains far from a single winner emerging through recursive self-improvement, given how close many models are in capability. However, the number of American laboratories capable of competing at the frontier may be narrowing to Anthropic, OpenAI, Meta, Gemini, and xAI.

Microsoft could join that group if it begins releasing models. The company has published strong scaling results and has the financial backing of a hyperscale business, along with direct access to OpenAI’s intellectual property for training models.

Models from smaller or mid-sized organizations such as Mistral, Nemotron, Reflection, and Arcee have found it increasingly difficult to influence either the open or closed frontier. They can still attract substantial adoption and support viable businesses, but developing a true frontier model now costs billions of dollars per major release. Only very large companies may be able to sustain that level of spending.

Despite the industry’s growing investment, its ability to predict real-world model success remains limited. Evaluations provide useful signals, but they do not fully explain why a laboratory can move from a successful model such as Claude Fable 5, to a weaker release such as Claude Opus 5, and then return to success with Opus 5.5. Accurately predicting these outcomes would be extremely valuable, yet the long-tail behavior of models remains difficult to anticipate.

From a long-term financial perspective, Meta, Gemini, and Microsoft may appear better positioned because of their strong balance sheets. This reflects the view that AI development will take time and that OpenAI and Anthropic may not possess an exclusive technological advantage. Given the industry’s rapid revenue growth, competition is likely to remain intense.

Meta’s valuation was below $2 trillion, while Anthropic was approaching its planned initial public offering at a similar valuation. Meta appeared approximately half a step behind on the same technology, but both companies could be viewed as starting from nearly the same position over the long term. Meta also has a highly profitable core business. This creates a substantial contrast between the two valuations.

As Meta and other competitors to OpenAI and Anthropic remain active, they may place downward pressure on valuations associated with AI initial public offerings. Frontier status can be valuable while also producing significant financial volatility. The eventual outcome remains uncertain.