Meta首款闭源大模型Muse Spark：Llama API运营仅14个月即下线 | Meta's First Closed-Source Model Muse Spark: Llama API Shut Down After Just 14 Months
Meta首款闭源大模型Muse Spark：Llama API运营仅14个月即下线
2025年5月，Meta上线Llama API，宣告”开源AI的基础设施层由我们提供”。14个月后，API下线，闭源模型Muse Spark登场。Meta的AI战略完成了一次180度转向。
2026年7月6日，Meta做了两件标志性的事：推出首款闭源大模型Muse Spark，同时正式下线Llama API公共预览服务。这两件事放在一起，勾勒出Meta AI战略的清晰转向。
Muse Spark：Meta的第一个闭源模型
Muse Spark是Meta推出的第一个闭源大模型。在此之前，Meta的AI品牌几乎等同于”开源”——Llama系列从Llama 2开始就是开放权重模型，是开源AI社区最重要的基石之一。
关于Muse Spark的技术细节目前披露有限，但从Meta的战略布局可以推断：
- 定位：面向商业API市场，直接与GPT-5.6和Claude Fable 5竞争。闭源意味着Meta可以控制模型的使用方式、收费模式和部署环境。
- 差异化：Meta的优势在于社交平台数据（Facebook、Instagram的图片和视频理解）和广告变现能力。Muse Spark很可能在多模态理解和商业场景理解上有独特优势。
- 部署方式：闭源模型不太可能开放权重下载，而是通过API或云端托管方式提供。这与Llama系列的”下载即用”模式完全不同。
Llama API：14个月的短命
Llama API于2025年5月上线，定位为”开源AI模型的基础设施层”——开发者不需要自己部署Llama模型，直接通过Meta的API调用即可。这是Meta第一次直接进入API服务市场，与OpenAI和Anthropic正面竞争。
但Llama API只运营了14个月就宣布下线。原因可能包括：
1. 商业回报不足：开源模型的API定价天然有天花板——用户可以自己部署Llama模型，API服务只能赚”便利性溢价”。如果便利性溢价无法覆盖服务器成本，API就是亏本生意。
2. 与闭源模型冲突：Muse Spark上线后，如果Meta同时维护开源Llama API和闭源Muse Spark API，会造成产品线内部竞争。下线Llama API是为了给Muse Spark让路。
3. 战略重心转移：Meta的CEO扎克伯格近期承认Meta的AI Agent开发”慢于预期”。与其在API基础设施上分散资源，不如集中力量做好模型本身。
“开源Llama + 闭源Muse”双轨制
Meta并非完全放弃开源。Llama系列模型仍然开放权重下载，开发者可以自行部署。但Meta不再提供官方API服务——想要用Llama，就得自己搭服务器。
这种”开源Llama + 闭源Muse”的双轨制意味着：
- Llama：面向有技术能力自行部署的开发者和研究机构。免费但需要投入基础设施成本。
- Muse Spark：面向需要即用型API的企业客户。付费但无需基础设施投入。
这种模式与Google的策略类似——Gemini是闭源API服务，而Gemma是开源权重模型。Meta正在从”开源唯一”走向”双轨并行”。
对开源社区的影响
Llama API的下线对开源AI社区是一个打击。许多中小型创业公司依赖Llama API作为低成本的模型推理服务——他们没有能力自行部署千亿参数模型。API下线后，这些公司面临三个选择：
- 迁移到其他开源API提供商：如Together AI、Fireworks AI等第三方Llama托管服务。成本可能更高，但至少模型权重是开放的。
- 转向闭源API：GPT-5.6、Claude、Muse Spark。性能可能更好，但失去了数据主权。
- 自行部署：需要GPU集群和技术团队。对小公司来说门槛很高。
Meta曾是开源AI最大的企业支持者。如果Meta的战略重心转向闭源，开源AI社区可能需要更多地依赖学术界和独立研究机构——而这些机构的算力资源远不及企业。
Meta’s First Closed-Source Model Muse Spark: Llama API Shut Down After Just 14 Months
In May 2025, Meta launched Llama API, declaring “the infrastructure layer for open-source AI is on us.” 14 months later, the API is shut down, and the closed-source Muse Spark takes the stage. Meta’s AI strategy has completed a 180-degree turn.
On July 6, 2026, Meta did two landmark things: launched its first closed-source large model Muse Spark, and officially shut down the Llama API public preview service. Together, these two moves outline a clear shift in Meta’s AI strategy.
Muse Spark: Meta’s First Closed-Source Model
Muse Spark is Meta’s first closed-source large model. Before this, Meta’s AI brand was virtually synonymous with “open source” — the Llama series has been open-weight since Llama 2, serving as one of the most important cornerstones of the open-source AI community.
Technical details about Muse Spark are currently limited, but from Meta’s strategic positioning we can infer:
- Positioning: Targeting the commercial API market, directly competing with GPT-5.6 and Claude Fable 5. Closed-source means Meta can control the model’s usage, pricing model, and deployment environment.
- Differentiation: Meta’s advantages lie in social platform data (Facebook and Instagram image and video understanding) and advertising monetization capabilities. Muse Spark likely has unique strengths in multimodal understanding and commercial scenario comprehension.
- Deployment: Closed-source models are unlikely to offer weight downloads, instead provided through API or cloud hosting. This is fundamentally different from Llama’s “download and use” model.
Llama API: 14 Months of Short Life
Llama API launched in May 2025, positioned as “the infrastructure layer for open-source AI models” — developers could call Llama models directly through Meta’s API without self-deployment. This was Meta’s first direct entry into the API services market, competing head-on with OpenAI and Anthropic.
But Llama API lasted only 14 months before being shut down. Reasons likely include:
1. Insufficient commercial returns: Open-source model API pricing has a natural ceiling — users can deploy Llama models themselves, so API services can only earn a “convenience premium.” If the premium can’t cover server costs, the API operates at a loss.
2. Conflict with closed-source models: After Muse Spark launches, maintaining both open-source Llama API and closed-source Muse Spark API would create internal product line competition. Shutting down Llama API clears the path for Muse Spark.
3. Strategic focus shift: Meta CEO Zuckerberg recently admitted Meta’s AI Agent development was “slower than expected.” Rather than spreading resources across API infrastructure, it’s better to concentrate on building the model itself.
“Open Llama + Closed Muse” Dual Track
Meta isn’t abandoning open source entirely. The Llama series remains open-weight for download — developers can self-deploy. But Meta no longer provides official API service — if you want to use Llama, you set up your own server.
This “open Llama + closed Muse” dual track means:
- Llama: For developers and research institutions with technical capability to self-deploy. Free but requires infrastructure investment.
- Muse Spark: For enterprise clients needing ready-to-use APIs. Paid but no infrastructure required.
This model mirrors Google’s strategy — Gemini is a closed-source API service, while Gemma is an open-weight model. Meta is moving from “open source only” to “dual track parallel.”
Impact on the Open-Source Community
Llama API’s shutdown is a blow to the open-source AI community. Many small and medium startups relied on Llama API as a low-cost model inference service — they lack the capability to self-deploy hundred-billion-parameter models. After the API shutdown, these companies face three choices:
- Migrate to other open-source API providers: Such as Together AI, Fireworks AI, and other third-party Llama hosting services. Costs may be higher, but at least the model weights are open.
- Switch to closed-source APIs: GPT-5.6, Claude, Muse Spark. Performance may be better, but data sovereignty is lost.
- Self-deploy: Requires GPU clusters and technical teams. High barrier for small companies.
Meta was once the largest corporate supporter of open-source AI. If Meta’s strategic focus shifts to closed-source, the open-source AI community may need to rely more on academia and independent research institutions — whose compute resources are far less than what corporations can provide.