Meta has unveiled Muse Spark Open, an open-weight AI model with 816 billion total parameters that activates 42 billion parameters per token. The model supports text, image, video and audio inputs, features a 1 million token context window, and targets coding and AI agent workflows. Meta describes it as the most capable US open-weight model to date, and developers can download it for customization.
In a notable hardware partnership, Meta and Apple teamed up to run Muse Spark Open on M5 Ultra-powered Mac Studios. Four such machines generate over 36 tokens per second, with each system offering up to 512GB of unified memory. The model requires roughly 493GB of memory in 4-bit format and must be split across machines linked by Thunderbolt cables. Apple says the M5 Ultra delivers up to 4.6 times faster time to first token than the M3 Ultra. The model is also compatible with SGLang, vLLM and Hugging Face Transformers.
The launch positions Meta against Chinese AI developers such as Alibaba's Moonshot AI and Z.ai, as competition in open-weight models intensifies globally.
On the financial side, AI-related debt concerns are rippling across Wall Street. Bonds funding data center projects are entering distressed territory, with rising yields reflecting investor worry about the financial sustainability of AI infrastructure spending.
Meanwhile, the hunger for training data is creating new market opportunities. AI companies are racing to acquire workplace data, and personal AI agents are driving large-scale data collection demand. This trend could revive a corner of Wall Street focused on data trading.
In other AI developments, Xbow USA researchers spent $40,000 in AI compute time using OpenAI's GPT-5.6 Sol to identify vulnerabilities in Google Chrome. The team chained flaws that could have compromised one billion Windows machines and escaped both the V8 JavaScript engine and operating-system sandboxes. Google patched the issues after Xbow disclosed its findings.
Malaysia and Indonesia are attracting billions in semiconductor and data center investments to capitalize on the AI boom. ASEAN foreign direct investment rose 8% to $226 billion in 2024, with Malaysia drawing $4 billion in data center investment and ranking among the top 10 globally. Both nations aim to move beyond low-value assembly work into chip design and research.
A WTW and MSCI Institute report found that 67% of global asset owners consider AI their top strategic priority. The study covered 20 asset owners managing over $5 trillion in assets, and participants expect technology's contribution to investment outcomes to grow from 17% to 30% within three to five years. However, 75% cite data quality as a key barrier, and only 25% have made AI a standing board agenda item.
McGraw Hill is expanding AI-powered study tools in its Connect platform through new Learning Studio capabilities, including AI-generated summaries, podcasts, knowledge checks and chapter overviews. The company also partnered with Breakout Learning, which reports that 90% of students using its modules finish assigned reading before sessions. The partnership gives instructors AI-powered feedback on critical thinking and communication at scale.
Key Takeaways
- Meta launched Muse Spark Open, an 816B parameter open-weight AI model activating 42B parameters per token with a 1 million token context window.
- Apple M5 Ultra Mac Studios run Muse Spark Open at over 36 tokens per second across four machines linked by Thunderbolt cables.
- Muse Spark Open is compatible with SGLang, vLLM and Hugging Face Transformers and positions Meta against Alibaba's Moonshot AI and Z.ai.
- AI data center bonds are showing signs of financial distress on Wall Street as rising yields reflect investor concern over AI infrastructure debt.
- AI companies are driving new demand for workplace data trading, creating fresh opportunities in financial markets.
- Xbow USA spent $40,000 using OpenAI's GPT-5.6 Sol to find Chrome vulnerabilities that could have affected one billion Windows machines.
- Malaysia attracted $4 billion in data center investment in 2024, and ASEAN FDI rose 8% to $226 billion despite a global decline.
- 67% of global asset owners making over $5 trillion in assets say AI is their top strategic priority, though 75% cite data quality barriers.
- McGraw Hill expanded AI study tools in Connect and partnered with Breakout Learning, where 90% of students finish reading before sessions.
- Poynter selected 33 journalists for its first Leading AI in the Newsroom program focusing on responsible AI adoption and editorial integrity.
Meta launches Muse Spark Open, an 816B parameter AI model
Meta Superintelligence Labs released Muse Spark Open, an open-weight AI model with 816 billion total parameters. It activates 42 billion parameters per token and supports text, image, video and audio inputs. The model features a 1 million token context window and is designed for coding and AI agent workflows. Meta claims it is the most capable US open-weight model to date.
Meta launches Muse Spark open model to challenge China's AI dominance
Meta launched Muse Spark Open on October 15, 2026, an 816 billion parameter open-weight AI model for complex coding and autonomous tasks. The model supports a 1 million token context window and can be downloaded by developers for customization. It runs on Apple M5 Ultra-powered Mac Studio systems and is compatible with multiple inference engines including SGLang, vLLM and Hugging Face Transformers. The launch positions Meta against Chinese AI developers like Alibaba's Moonshot AI and Z.ai.
Meta and Apple partner to run 816B parameter AI model on Mac Studios
Meta and Apple partnered to run Muse Spark Open on M5 Ultra Mac Studios. Four M5 Ultra Mac Studios generate over 36 tokens per second with the model. Each machine provides up to 512GB of unified memory, and the model requires about 493GB of memory in 4-bit format. The model must be split across machines linked by Thunderbolt cables. Apple said the M5 Ultra delivers up to 4.6 times faster time to first token than the previous M3 Ultra.
AI debt fears ripple across Wall Street as bond yields surge
AI debt concerns are spreading across Wall Street as bonds funding data center projects enter distressed territory. More bonds are showing signs of financial distress. The rising bond yields reflect growing investor worry about AI-related debt. The situation highlights financial risks in the rapidly expanding AI infrastructure sector.
AI's need for data could revive a corner of Wall Street
AI companies are racing to acquire workplace data to train their models. Personal AI agents are driving demand for large-scale data collection. This trend could bring new business to a previously active area of Wall Street focused on data trading. The surge in AI data needs is creating fresh opportunities in financial markets.
ActiveEvidence uses decision-focused evidence graphs for AI
ActiveEvidence introduces a method called decision-certifiable active evidence acquisition for multimodal AI agents. It uses a Counterfactual Evidence Interaction Graph to evaluate whether visual evidence is worth collecting. The system achieved a task score of 0.842 and a p95 latency of 742ms across 640 test cases. Dense replay of certificates showed 0.992 action agreement while evaluating fewer expensive interaction terms.
Malaysia and Indonesia chase AI investment and tech growth
Malaysia and Indonesia are attracting billions in semiconductor and data center investments to gain from the AI boom. ASEAN foreign direct investment rose 8% to $226 billion in 2024 despite a global decline. Malaysia attracted $4 billion in data center investment, ranking among the top 10 globally. Both countries aim to move beyond low-value assembly work into higher-tech activities like chip design and research.
Meta AI Reels show misleading content with crying children
AI-generated content on Facebook and Instagram is showing distressing clips of crying children with misleading messages. One clip features a sobbing child clutching bread and tells viewers they will be freed if people like and share the video. The content raises concerns about harmful AI-generated material spreading on Meta platforms.
Xbow spent $40K on AI to crack Google Chrome
Researchers at Xbow USA spent $40,000 in AI compute time using OpenAI's GPT-5.6 Sol to find vulnerabilities in Google Chrome. They chained flaws that could have allowed control of one billion Windows machines worldwide. The team escaped both the V8 JavaScript engine sandbox and Chrome's operating-system sandbox. Google fixed the flaws after Xbow presented its findings.
67% of global asset owners make AI top priority
A WTW and MSCI Institute report found 67% of global asset owners say AI is their top strategic priority. The study covered 20 asset owners managing over $5 trillion in assets. Participants expect technology's contribution to investment outcomes to grow from 17% to 30% in three to five years. Only 25% have made AI a standing board agenda item, and 75% cite data quality as a key barrier.
PROBE trains AI auditors to find policy violations in multi-turn agent tasks
PROBE is a new algorithm that trains AI auditors to test other AI agents for policy violations during real tool use. The auditors interact with agents as realistic users and probe them through multi-turn dialogue. The method formalizes auditing as a two-player Markov game and compares training approaches like GRPO, DPO, and GFlowNet. Results show PROBE auditors find broader policy violations than smaller zero-shot attackers.
Poynter selects 33 journalists for first AI leadership newsroom cohort
Poynter announced 33 journalists and newsroom leaders chosen for its first Leading AI in the Newsroom program. The cohort will focus on responsible AI adoption, editorial integrity, and building trust among colleagues. Participants will attend an in-person workshop in Los Angeles in late October, followed by four virtual sessions through early December. Leaders including Alex Mahadevan, Tony Elkins, Kelly McBride, and Sitara Nieves will guide the program, with guest insights from Benjamin Toff and Christine DiGangi.
McGraw Hill expands AI study tools and partners with Breakout Learning
McGraw Hill is expanding AI-powered study features in its Connect platform and partnering with Breakout Learning for student discussion tools. New Learning Studio capabilities offer AI-generated summaries, podcasts, knowledge checks, and chapter overviews tied to course content. Breakout Learning reports that 90% of students using its modules finish assigned reading before sessions. The partnership gives instructors AI-powered feedback on critical thinking and communication at scale.
Sources
- Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context
- Meta launches Muse Spark open model to challenge China’s AI dominance
- Meta and Apple are partnering to run an open 816B-parameter AI model on Mac Studios
- AI debt fears are suddenly rippling across Wall Street as yields surge
- AI's need for data could breathe new life into a once-buzzy corner of Wall Street
- ActiveEvidence: Decision-Focused Evidence Graphs with Verifiable Acquisition Gaps
- Malaysia and Indonesia pursue tech ambitions amid AI investment surge
- Meta's AI Reels show crying kids 'arrested' and told they'll be freed if viewers like and follow
- Xbow spent $40K on AI and cracked Google Chrome wide open
- Report finds 67% of global asset owners cite AI as top strategic priority
- PROBE: Learning to Audit Policy Compliance in Tool-Using LLM Agents
- Poynter announces first cohort for Leading AI in the Newsroom
- Breakout Learning reports 90% of module users finish assigned reading before sessions. McGraw Hill is partnering with it.
Comments
Please log in to post a comment.