Facebook and Instagram parent Meta Platforms, Inc. (META.US) released Muse Spark Open on October 15, fully opening the weights of this large model focused on AI supercoding and AI agent tasks, while keeping the more advanced Muse Wonder, launched earlier this week with far superior overall performance metrics, closed-source.
This latest large-model move signals that the company's commercial logic at the AI application and large-model level is gradually becoming clear — using open-weight models to attract more developers to the Meta community and expand computing resource demand and AI application deployment, while also focusing on high-end models, developer tools and more powerful inference efficiency in cloud computing workloads to capture more expensive paid AI subscription demand.
Meta stated that Muse Spark Open is "the most capable open-weight AI large model in the United States to date," so from the perspective of expanding AI computing demand, this open-source move is a major positive catalyst for the semiconductor and AI computing theme that has been plummeting recently — from local workstations to enterprise private AI computing clusters to public clouds, open weights are expected to bring agents into more scenarios previously difficult to cover due to cost, data control and customization needs, providing new momentum for the continued massive expansion of computing demand where "closed-source and open-source AI flourish together."
In the AI developer ecosystem, open-weight AI typically refers to "open-weight AI large models," meaning developers can download the trained model parameters and deploy, fine-tune and run inference on their own servers or cloud environments; however, the model developer may not disclose training data, data processing methods, training code, or complete architecture details, and may restrict commercial use or specific applications through licenses. Open-source AI in the strict sense has a broader scope: in addition to weights, it should provide code, architecture and training data information sufficient for users to study, modify and reproduce the system, and guarantee basic freedoms of use, research, modification and redistribution. Therefore, open weights mean "the finished model can be used and modified," while true open-source AI further opens up "how the model was made"; the former can be part of the latter, and in current common usage, "open source" usually refers to "open-weight AI large models," but open weights do not necessarily equal open-sourcing all technical components and code.
Meta's blockbuster release comes at a time when chip stocks are under pressure. As of the October 15 U.S. stock market close, the Nasdaq 100 fell 0.05%, down 3.07% for the week, down 1.56% for October, but still up 18.56% year-to-date; it has pulled back 4.13% from this month's closing record, not yet entering a technical correction typically defined by a 10% decline. The Philadelphia Semiconductor Index, known as the "barometer of AI computing theme trading," fell 0.88%, marking a seventh consecutive decline with a cumulative drawdown of 12.38% entering technical correction territory, and down 8.29% for October. The market is currently reassessing the increasingly high global financing costs and unclear AI capital return cycles against the backdrop of resurgent high inflation, surging 10-year Treasury yields that remain at twenty-year highs. Long-end risk-free rates remain elevated, driving further expansion of credit risk premiums, which together raise financing costs — this is the core mechanism suppressing AI computing themes and semiconductor stocks. On October 15, longer-dated U.S. Treasury yields of 10 years and above retreated somewhat, combined with Jensen Huang's positive comments on CoreWeave's outlook that eased selling pressure, but semiconductors still seem unable to stop falling. Therefore, the potential investment significance of Meta's latest open model lies in whether it can expand payable AI application demand, improve computing utilization, and ultimately translate purchases of core AI infrastructure hardware such as AI chips, memory chips and data center CPUs into more solid cash returns.
Meta Opens Its Model While Retaining a Closed-Source Business Model for Its Most Advanced AI Technology
Facebook and Instagram parent Meta Platforms, Inc. announced on Thursday local time that it is opening a new artificial intelligence model to developers worldwide, allowing downloads and modifications, while continuing to keep its most advanced technology closed-source in an effort to convert AI investments into revenue. On Thursday, Meta announced the release of the "open weights" for Muse Spark Open — allowing developers to run and adjust the model's parameter settings themselves. Meta stated that this model, designed to support coding and AI agents, is less advanced than Muse Wonder launched earlier this week. Releasing this relatively less capable open-weight model fulfills Meta's previous commitment to making its technology widely available. "For both safety and capability reasons, we typically don't open-source our most powerful models," Meta Chief AI Officer Alexander Wang said in an interview. "But we do believe it's very important to continue open-sourcing excellent models, and we intend to keep doing so." Leading Chinese AI large-model companies including Alibaba Group and DeepSeek have achieved tremendous success in developing open-source AI models that allow developers to adjust them. These models were rapidly adopted, including in the United States, further intensifying technological competition between the two countries. Meta CEO Mark Zuckerberg has long advocated opening AI models and their building blocks to the developer community. But last year, after the company's open-weight Llama models disappointed Silicon Valley, Zuckerberg restructured Meta's AI division and brought in Wang. Since then, Wang has helped develop a series of more capable but closed-source models. Wang said he believes it is important for American companies to release open-weight models, helping them maintain their lead over Chinese AI large-model leaders in the rapidly evolving AI race. "Most AI in the world will run on open models," he said. "These models should come from the United States." At the same time, he expects Chinese AI developers to study Muse Spark Open and continue building on it. "If the Chinese model developer community incorporates some of our technical principles into their models, that might even be a good thing," he said. The release of Muse Spark Open comes as Meta is also building its own cloud business, selling access to AI models and computing infrastructure through it. Wang said offering Muse Spark Open for free would not undermine this business. Instead, he views it as a customer acquisition funnel: developers can experiment on their own first, and when they need more computing power, they pay Meta for additional tools. Developers can currently access Muse Spark Open through platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Model weights can be downloaded on Amazon SageMaker AI, Google Cloud Model Garden and Hugging Face. Meta stated it partnered with Apple to ensure the model can run on two Mac Studio desktop computers equipped with M5 Ultra chips, allowing developers to use their own hardware without relying on remote servers.
Open Models Broaden Applications, While Closed-Source Models Like Astra Are Expected to Keep Raising the Capability Ceiling — Closed Source Competes for the Capability Ceiling, Open Source Broadens and Strengthens Computing Demand Boundaries
Notably, "open weights" are a delivery method, not a source of performance; openness changes how developers obtain, deploy and modify models. "Open weights" mean developers can obtain trained model parameters and deploy, fine-tune and adapt them to their business themselves. Its industrial value lies in lowering the barrier to application deployment, expanding models from services provided by a few platforms into computing capabilities that more enterprises can use independently, thereby opening broader growth space for inference calls and computing demand. Muse Spark Open's differentiated advantages are mainly concentrated in the combination of efficient tool execution related to AI agents, native multimodality and self-deployability. It uses a mixture-of-experts (MoE) architecture with 816 billion total parameters and approximately 42 billion activated parameters per token, supports a 1 million token context, accepts text, image, audio and video inputs, and outputs text. In Meta's comparison published by model with specified inference tiers, its MCP-Atlas tool task score was 89.2, higher than GLM 5.3's 84.2 and Kimi K3's 82.3; but its overall intelligence index was 39, lower than the other two's 45 and 44, and its Terminal-Bench 2.1 score was 76.4, also lower than 83.9 and 85. Therefore, Meta's latest open-source AI large model Muse Spark Open has competitive advantages on some agent tasks, but its overall positioning remains below the closed-source Muse Wonder and cannot be summarized as comprehensively leading mainstream models. The Apple partnership further expands deployment options: with MLX and quantized versions, two M5 Ultra Mac Studio computers each with 512GB unified memory can run the model, providing a path for private data processing and local research and development. From a demand mechanism perspective, closed-source frontier models expand the set of tasks AI can accomplish, while open models lower the deployment barrier for these capabilities to enter actual business operations. Enterprises can conduct domain fine-tuning, agent development and private deployment without redoing expensive foundational pre-training; when the cost per effective task decreases, work previously not worth automating may also gain commercial value. The key to whether the so-called "Jevons Paradox" fully holds in the current market discussion is whether the volume of new tasks exceeds the savings in computing power per task. Performance improvements alone do not guarantee total computing growth, but the simultaneous emergence of stronger capabilities, lower costs and broader deployment significantly increases the likelihood of demand expansion. The so-called Jevons Effect, also known as the Jevons Paradox, is a counterintuitive economic theory: when frontier technological progress improves the efficiency of using a resource (such as energy, raw materials or AI computing infrastructure resources), it lowers unit costs, thereby stimulating large-scale market demand expansion, ultimately causing total consumption of all types of resources to increase rather than decrease. The concept was first proposed by British economist William Stanley Jevons in 1865 in his book "The Coal Question." For global investors focused on the AI computing industry chain and the AI super bull market investment direction, open weights, model routing and architectural efficiency together significantly reduce unit intelligence costs, yet may amplify total computing demand through the "Jevons Paradox." When simple invocation costs decline, enterprises will deploy more continuously running agents, parallel sub-agents, long-context analysis, code automation and real-time multimodal services; a single task may consume less computing power, but the number of tasks, inference steps and deployment nodes may grow faster. Thus, the large-scale accelerated adoption of open-weight large models will spread computing demand from a few closed-source labs to cloud computing providers, sovereign clouds, giant AI data centers for enterprises and local inference clusters worldwide.
Stock Prices Pull Back, but the Growth Logic Still Expands! As Meta Broadens Application Boundaries, Is a Semiconductor Counterattack Brewing?
Recent research by Google parent Alphabet explicitly points out that lower unit token prices, longer task execution times and increased application numbers can occur simultaneously, driving up total AI inference demand and total inference value — precisely corresponding to the so-called "Jevons Paradox." A forecast report from well-known market research firm TrendForce shows that 2027 NVL72 rack shipments will grow more than 50% year-over-year, and the total output value of NVL rack systems based on the GB200/GB300 and next-generation Vera Rubin architecture surrounding Nvidia AI GPUs is expected to exceed $710 billion, up 214% year-over-year, including the impact of product upgrades and price increases. The most core change brought by the recently globally popular Meta Muse AI agent and the "AGI-comparable" OpenAI Astra large model/AI agent is that a single user instruction can trigger sustained, multi-stage computational work. A research, coding or office task may sequentially include planning, retrieval, document reading, tool invocation, code execution, result checking and error correction, with multiple steps requiring model re-invocation, and complex tasks may also adopt parallel exploration and verification — and such nearly endless and increasingly complex AI workloads will accelerate growth opportunities to the complete AI inference-focused computing system beyond GPUs. Meta's opening of Muse Spark Open provides new technological momentum for application diffusion and inference demand growth. As Meta Muse and OpenAI's GPT-6 Astra fully ignite the AI agent wave, combined with the latest development showing the release of Muse Spark Open — "the most capable open-weight AI large model in the United States to date" — also focused on cutting-edge AI agent workloads, "AI FOMO trading" (referring to investors' fear of missing out on gains in AI-related assets, thus chasing or replenishing previously reduced positions) is likely to make a comeback in the near term. From the perspective of the underlying architecture of AI computing infrastructure, Muse and Astra, as well as Meta's latest open-source Muse Spark Open, are expected to jointly further expand the scope of AI agent usage; the massive expansion of AI inference workloads they bring will simultaneously increase demand for model inference, tool execution and state management, while deepening competitive pressure on some software companies. Accelerators such as GPUs and TPUs handle model computation; high-performance server CPUs run browsers, virtual machines, product searches, database queries and transaction orchestration; HBM and server DRAM carry model data, context and concurrent working environments; enterprise SSDs store product indexes, task records and persistent states, while high-speed networking and optical interconnects support distributed data exchange. Wall Street financial giant JPMorgan recently stated that its global equity market strategist team favors a repositioning opportunity formed by easing position crowding, falling valuations and earnings resilience, with a particular preference for semiconductors. According to the institution's latest research report, since June, semiconductor forward 12-month EPS forecasts have been raised by about 30% to 40%, while global software sector earnings expectations lack corresponding improvement; the institution's latest cited capital expenditure outlook for hyperscale cloud providers is approximately $950 billion for 2026, about $1.4 trillion for 2027, and at least about $3 trillion for 2030, and it expects AI-related revenue growth to show explosive expansion in 2027. Goldman Sachs technology industry expert Peter Callahan also noted in a recent research view that the long-term AI investment trend remains intact, semiconductor valuations have already compressed, and the development of personal assistants and consumer-grade AI is expected to help the sector regain its footing during the third-quarter earnings season. The latest research reports from Goldman Sachs and JPMorgan support the bullish judgment that "AI computing and semiconductors have strong conditions for a counterattack," and respectively fill in key evidence on the funding side and earnings side. Goldman Sachs focuses on when incremental buyers will return, while another JPMorgan report on the bullish outlook for a South Korean stock market rebound judges whether the high memory chip earnings crucial to AI computing trades can persist longer than the market expects. The "three major reinforcements" described by Goldman Sachs are expected to improve the funding environment for AI assets to regain pricing. Goldman Sachs' capital flows team noted in its latest report "Witching Hour" that, according to its estimates, about 48% of S&P 500 constituents will re-enter the buyback open window by the end of October; historically, November contributes about 10.1% of full-year buyback execution volume, and combined with the fading of mutual fund fiscal-year-end tax selling pressure, year-end retail inflows, and overseas allocation demand after the midterm elections, stock supply-demand relations may gradually improve. Among these, about $190 billion in overseas buying is a potential scale extrapolated from historical patterns. For semiconductors that have experienced concentrated de-risking, a recovery in buying means that even if companies merely prove orders and earnings have not deteriorated significantly, it could trigger position replenishment. However, tech sector buyback execution volume fell 40% year-over-year, indicating that capital expenditure is competing with shareholder returns for funds; for AI stocks to continue attracting incremental allocation, they still need to prove that capital investment in key AI infrastructure resources such as GPUs, network infrastructure, data center optical interconnects and CPUs can generate higher future cash flows. JPMorgan's bullish logic falls more directly on "whether the market is prematurely discounting a decline in memory earnings." A semiconductor research report on the South Korean stock market published by the institution on October 15 shows that macro-level concerns about the sustainability of AI investment are clearly disconnected from bottom-up supply and demand data in the memory industry, so it views the third quarter as an "expectation reset quarter," and over the next six months will core overweight Samsung Electronics, SK Hynix and Kioxia, extending to Samsung Electro-Mechanics, AI PCBs and test sockets. The "duration trade" here bets that the upcycle lasts longer and high earnings persist longer: even if the pace of price increases slows, as long as HBM, server DRAM and enterprise SSD demand continues to absorb new capacity, the rapid profit decline previously priced in by the market may need to be revised. This also explains why memory stocks can fall before earnings peak, yet may rebound early once earnings duration is verified; what determines the opportunity is the gap between the actual earnings path and the expectations implied by stock prices. Meta's open model adds application-side evidence for "earnings duration may be underestimated." Closed-source frontier models such as Muse Wonder expand the task boundaries agents can complete, while open-weight models like Muse Spark Open lower the barriers to enterprise customization, private deployment and migration, making more workflows economically feasible. From the underlying AI infrastructure technical mechanism, a complete task in the most cutting-edge AI agent task system may include repeated planning, code execution, retrieval, tool invocation and result verification, consuming not only GPU/TPU/ASIC computation but also CPU execution resources, KV cache, host memory, SSD and high-speed interconnects. MoE sparse activation reduces computation per token, but it does not mean that full model weight storage and communication demand disappear proportionally. From this, when the growth in new task volume, concurrency and workflow length exceeds the decline in computing consumption per task, the Jevons effect may drive total computing demand expansion; the significance of coordinated development of open and closed source lies precisely in simultaneously raising the capability ceiling and expanding application coverage.
Comments