model release

Meta launches proprietary Muse Spark model, abandoning open-source commitment

TL;DR

Meta has released Muse Spark, a proprietary AI model with restricted access via API and portal invite only—a striking reversal from CEO Mark Zuckerberg's 2024 manifesto championing open-source AI. The model claims performance matching top competitors from OpenAI, Anthropic, and Google, trained with an order of magnitude less compute than Llama 4.

3 min read
0

Meta Launches Proprietary Muse Spark Model, Reversing Open-Source Stance

Meta has unveiled Muse Spark, a closed proprietary AI model developed by its Superintelligence team, marking a significant departure from the company's previous commitment to open-source AI development.

The model is available exclusively through Meta's AI portal and via API access on an invite-only basis. Unlike Llama, Meta's previous open-weight models, Muse Spark's weights are not publicly available for download.

Direct Contradiction to Earlier Commitment

This move directly contradicts statements made by CEO Mark Zuckerberg in his 2024 manifesto titled "Open Source AI is the Path Forward." In that document, Zuckerberg argued that "open source AI represents the world's best shot at harnessing this technology to create the greatest economic opportunity and security for everyone." He compared open-source AI to Linux, claiming Meta's business model didn't depend on selling access to models.

Yet just over a year later, Meta launched an inference API for Llama 4, demonstrating a shift toward monetized access. Now, with Muse Spark entirely proprietary, the pattern is clear.

Performance Claims and Benchmarks

According to Meta, Muse Spark matches or exceeds performance from OpenAI, Anthropic, and Google on standard benchmarks. The company claims the model was trained with "an order of magnitude less compute than our previous model," referring to Llama 4.

The benchmarks presented by Meta should be interpreted cautiously—the company previously faced accusations of inflating Llama 4's performance metrics. However, Meta states it is sharing test methodology this time, providing some transparency into how claims were validated.

Architecture and Capabilities

Meta describes Muse Spark as a "natively multimodal reasoning model" with tool-use, visual chain-of-thought, and multi-agent orchestration capabilities. The model includes a "contemplating mode" that orchestrates multiple reasoning agents in parallel to compete with frontier models like Google's Gemini Deep Think and OpenAI's GPT Pro. This mode is rolling out gradually rather than being available at launch.

Context of Previous Failure

The shift toward a proprietary model follows Llama 4's underperformance relative to hype. Meta had been developing "Behemoth," a 2-trillion-parameter variant intended as its flagship model, but abandoned it midway through development. The failure was reportedly significant enough to prompt a complete organizational restructuring—Meta hired high-profile AI talent including Alexandr Wang to lead the newly formed Superintelligence Labs. This reset appears to have resulted in Muse Spark as the first model from the reorganized effort.

Mixed Future Commitments

Zuckerberg stated in a Threads post that Meta plans to "release increasingly advanced models that push the frontier of intelligence and capabilities, including new open source models." This suggests Muse Spark is not the end of Meta's open-source work, but rather a dual approach: proprietary models for monetization and select open models for ecosystem development.

Google and OpenAI have adopted similar dual-track strategies, releasing smaller open-weight models alongside larger proprietary ones. However, the contradiction between Zuckerberg's 2024 manifesto and current practice raises questions about whether this was always the plan or a strategic pivot driven by competitive pressure.

What This Means

Meta is openly abandoning its open-source purist position in favor of a more pragmatic commercial model. Muse Spark represents a $135 billion bet that frontier performance requires proprietary control—at least in the near term. The success of this model will determine whether Meta's AI narrative can shift without further damage to its credibility. For users, the practical impact is clear: access to Meta's best models now requires API payments or invitation, fundamentally changing the company's competitive positioning from democratizer to gatekeeper.

Related Articles

model release

Mistral Large 4 enters public preview: 1T-parameter open-weight multimodal model, weights due by end of October

Mistral AI has launched a public preview of Mistral Large 4, a 1-trillion-parameter natively multimodal model with 49 billion active parameters. The preview API is live on Mistral Studio, and open weights are promised by the end of October 2026. Pricing and context window have not been disclosed.

model release

Google DeepMind releases EmbeddingGemma 2: 740M-parameter open embedding model spanning text, image, video, audio

Google DeepMind has released EmbeddingGemma 2, an Apache 2.0 open embedding model with 740M total parameters that maps text, images, video and audio into a single 768-dimensional vector space. According to the model card, it improves code retrieval on MTEB (code, v1) from 68.76 to 78.68 over its predecessor while keeping an 8,192-token context window.

model release

Google releases EmbeddingGemma 2: 740M-parameter multimodal embedding model under Apache 2.0

Google announced EmbeddingGemma 2, a 740M-parameter natively multimodal embedding model built on the Gemma 4 architecture and released under Apache 2.0. Google says it runs in ~191MB of active RAM for text-only weights and ~567MB for the full multimodal model on a quantized Pixel 11 Pro. Google also released a Mac app, AI Edge Foresight, to demonstrate it.

model release

Google releases EmbeddingGemma 2: 740M-parameter multimodal embedding model under Apache 2.0

Google announced EmbeddingGemma 2, a 740M-parameter natively multimodal embedding model built on the Gemma 4 architecture and released under Apache 2.0. Google says the quantized model needs about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model on a Pixel 11 Pro. Google also launched a Mac app, AI Edge Foresight, to demonstrate it.

Comments

Loading...