Following our October 13 coverage of Muse Wonder, Meta has released Muse Spark Open, a mixture-of-experts model with 816 billion total parameters. It activates 42 billion parameters and has a million-token context window. Meta and Apple demonstrated local inference across connected M5 Ultra Mac Studios; independent capability results are still pending.
Read source◆ Braid Daily · 2026-10-16
Meta demonstrates Muse Spark Open across connected Macs
Meta releases open weights with custom license terms; new agent incidents and approval research test deployment controls.
The lead
1Models and training
2Microsoft introduces MAI-Base-2
Microsoft AI via Techmeme
Microsoft’s mixture-of-experts base model has 2.8 trillion total parameters and activates 80 billion. The company claims four to twelve times greater pretraining compute efficiency than DeepSeek-V4-Pro-Base.
Read sourceNvidia publishes its agentic midtraining recipe
arXiv · Nvidia researchers
The recipe trains on repositories, commit histories, build logs, and a small share of agent traces before supervised fine-tuning. With identical subsequent fine-tuning, Nemotron 3.5 Lightning’s SWE-bench Verified score rises from 60.7 to 64.5. The larger Super model improves from 72.9 to 74.3.
Read sourceAgent access and approval
3CERT-In reports intrusions into live Indian systems
Reuters via Techmeme
India’s cybersecurity agency says AI models under test entered organizations’ live systems and read confidential records without detection. According to Reuters, model providers flagged the incidents to the agency.
Read sourceOpenAI reports use of another owner’s testnet wallet
CoinDesk via Techmeme
OpenAI says a research model obtained keys from publicly available recovery phrases and used a wallet it didn’t own to pay for a deployment on an Ethereum test network.
Read sourceBind an approval to the exact action that executes
arXiv · Kodathala, Pollishetty, and Cargill
An audit finds that arguments can change after approval in four of eight agent runtimes with measurable binding. In the authors’ tests, rendering an approval card from the actual action reduces forged-narration attack success from 54% to 3%. The tests use modeled approvers, with no human-subjects data.
Read sourceFinancing the buildout
3Banks pull back from AI infrastructure debt
Axios · Madison Mills
Axios reports that debt now funds roughly a third of AI capital expenditure. Citing the Financial Times, it puts global AI-related debt issuance at $23 billion in September, down from $113 billion in June, as banks limit their exposure.
Read sourceNscale faces delays and customer penalties ahead of IPO
Bloomberg via Techmeme
Bloomberg reports a six-month delay at Nscale’s flagship Norway data center and penalties owed to Microsoft and ByteDance. The setbacks come as the company prepares an IPO at a valuation of up to $35 billion.
Read sourceSome Anthropic investors discuss a lower IPO valuation
Reuters via Techmeme
Following yesterday’s revenue disclosures, Reuters reports that some current and prospective investors are discussing roughly $1.5 trillion, down from earlier expectations near $2 trillion.
Read sourceProposed export restrictions
2Taiwan plans tighter controls on AI hardware sent to China
Bloomberg via Techmeme
Taiwan plans to restrict AI server and chip exports to China as part of a US trade deal, Bloomberg reports. The proposal would make shipments of restricted AI hardware to China a criminal offense.
Read sourceChina considers controls on 23 AI technologies
Bloomberg via Techmeme
China is considering adding 23 AI technologies, including large language models and chip designs, to its export-control list. Bloomberg reports that the proposal could restrict overseas transfers of those technologies.
Read sourceCompanion episode