Meta Launches Native Multimodal Large Model Muse Spark
Meta Superintelligence Labs (MSL) launches its first model, Muse Spark. On Artificial Analysis benchmarks, scores jumped from 18 for Llama 4 to 52, ranking second only to GPT-5.4 and Gemini 3.1 Pro, driving Meta's stock price up nearly 10%. The model features a native multimodal reasoning architecture with visual chain of thought, multi-Agent orchestration, and a "contemplation mode," performing notably well in CharXiv visual understanding and HealthBench health Q&A domains.
Meta is accelerating its layout in the artificial intelligence field. On July 9 local time, Meta released the latest version of its AI model, Muse Spark 1.1, focusing on the AI coding and intelligent agent (AI Agent) markets, attempting to catch up with OpenAI and Anthropic, who currently lead in the AI field.
Meta AI head Alexandr Wang stated that Muse Spark 1.1 is Meta's "most powerful intelligent agent and coding model."
Meta opens Muse Spark API, expanding developer usage scope
In April this year, Meta first launched the Muse Spark model, but at that time, it was only open to a few partners via private API testing methods.
With the launch of Muse Spark 1.1, Meta began opening API interfaces through the developer portal, allowing more developers to apply for access.
However, Meta currently still restricts the model to run on its own computing infrastructure and has not opened it to third-party AI model platforms.
Meta stated that some early partners already have access to the API, and new users can join the waitlist to gradually obtain usage rights in the future.
Alexandr Wang stated:
"This model will run on the computing infrastructure we have already built."
AI coding becomes a key competition point
Muse Spark 1.1 primarily strengthens AI coding capabilities.
Meta stated that the model has been optimized for handling code development, calling development tools, and executing complex tasks.
With the rapid development of AI Agent technology, coding capability is considered an important foundation for building the next generation of intelligent assistants.
AI Agents can autonomously complete multi-step tasks like "digital employees," such as:
writing and modifying code;
debugging software;
using third-party tools;
analyzing data;
executing complex workflows.
Alexandr Wang stated that the Meta Superintelligence Labs team specifically trained Muse Spark 1.1 to perform stronger in coding-related tasks, because powerful coding capability is an important component of realizing advanced AI Agents.
Meta adopts a low-price strategy to attract developers
To compete with OpenAI and Anthropic, Meta launched competitive pricing.
Alexandr Wang stated that the price of Muse Spark 1.1 is "very aggressive and attractive" compared to similar AI models on the market.
After registering an API account, new users will receive a $20 free credit.
Subsequent charging standards are:
Input content: $1.25 per million tokens;
Output content: $4.25 per million tokens.
Meta hopes to attract a large number of developers to use its AI models through lower prices.
Wang stated:
"Our goal is to truly provide attractive pricing, allowing the model to continuously expand with large-scale usage."
Meta shifts from open-source route to commercial AI services
Over the past few years, Meta has primarily promoted the open-source AI ecosystem through the Llama series models.
However, as AI industry competition enters the commercialization stage, Meta is attempting to generate revenue by selling access to proprietary models.
Alexandr Wang stated that Meta remains "committed to open source" and plans to launch a version of Muse Spark in the future and open its source code to the community.
However, he did not reveal a specific release time.
Zuckerberg faces AI investment return pressure
Meta CEO Mark Zuckerberg is currently facing pressure from Wall Street.
In recent years, Meta has invested heavily in building AI infrastructure, including:
large-scale data centers;
AI computing equipment;
model R&D teams.
However, compared to OpenAI, Anthropic, and Google, Meta is still in the catching-up stage regarding popular AI products and commercial applications.
Although Meta possesses strong computing resources, it currently does not have a mature cloud computing business, and how AI investments convert into revenue remains a concern for investors.
Meta explores AI health assistant applications
Alexandr Wang revealed that he is personally testing Muse Spark 1.1 and exploring AI Agent applications in the personal health field.
For example:
searching medical materials;
reading research papers;
analyzing personal health data.
He believes that the health field is one of the application scenarios that best reflects the value of intelligent agents.
Stronger AI model Watermelon is under development
In addition to the Muse Spark series, Meta is also training a more powerful AI model, internally codenamed Watermelon.
Meta has not yet announced the release time for this model.
Previously, the internal codename for Muse Spark was Avocado.
Closing remarks
Meta is adjusting its AI strategy.
In the past, Meta relied mainly on open-source Llama models to expand influence; now, facing the leading advantage of OpenAI and Anthropic in the commercial AI market, Meta has begun launching paid API services to directly participate in the foundation model competition.
The launch of Muse Spark 1.1 marks Meta's formal entry into the AI coding and intelligent agent market.
The key to future AI competition is not only model capability but also who can attract more developers and enterprises to truly use their technology.
This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.