Meta Muse Spark Open Runs 816B Params on Mac Studios
Meta released Muse Spark Open, an 816 billion parameter mixture of experts model with 42 billion active parameters and a 1 million token context window. It ships under a new Muse Model License with day-one support in SGLang, vLLM and TensorRT-LLM, and runs on two to four paired M5 Ultra Mac Studios through MLX. Nvidia released Nemotron 3.5 Super VL the same day, a 120 billion parameter hybrid Mamba-Transformer with 12 billion active parameters under the OpenMDW licence, claiming 5 times the throughput of comparable models and wins on five tasks against Claude Opus 5 and GPT-6.1 Sol.
The regulatory side moved just as fast. The Federal Trade Commission sent 50-page civil investigative demands to OpenAI and Anthropic one day after the European Commission opened its first AI Act inquiries into the same two companies, eight state attorneys general sued OpenAI over alleged harm to minors, and the White House adopted a charter for a Super Intelligence task force with a 120-day reporting window. Here are the 16 updates that matter most. The AI industry news and trends hub carries the running archive.
What Is Meta Muse Spark Open? An 816B Mixture of Experts
Muse Spark Open is Meta's new open weights flagship: 816 billion total parameters, 42 billion active per token, and a 1 million token context window, released under a bespoke Muse Model License. Serving support arrived on day one across SGLang, vLLM and TensorRT-LLM rather than weeks later, which is a change from how most large open releases have landed this year.
The architecture choice is the interesting part. A 19 to 1 ratio between total and active parameters is unusually sparse, and it is what makes the local-hardware story possible: you need enough memory to hold 816 billion parameters but only enough compute to run 42 billion of them per token. Memory is cheap to add in unified-memory machines and compute is not, so the design is tuned for exactly the hardware it is being demonstrated on.
Why this matters: the open weights frontier has been stuck in a band between 100 and 400 billion parameters for most of the year, with Tencent's 770 billion parameter Apache 2.0 release in late August the main exception. An 816 billion parameter model with a million-token window, released by Meta with real serving support, resets what the open ceiling looks like. Read the licence before building on it, since Meta called it a Muse Model License rather than using Apache or MIT, and bespoke model licences have historically carried use restrictions worth checking against your product.
Can You Run Muse Spark Open Locally? Two Mac Studios Will Do
Meta says Muse Spark Open runs on two to four paired M5 Ultra Mac Studios through MLX, Apple's array framework for Apple silicon. Pairing Mac Studios over Thunderbolt to pool unified memory has been a hobbyist technique for over a year; what is new is a frontier-scale open model whose official deployment guidance names it as a supported configuration.
The economics are worth doing properly before anyone gets excited. Four high-memory M5 Ultra Mac Studios is a serious capital purchase, but it is a one-time cost with no per-token billing, no rate limits and no data leaving your premises, and it sits on a desk rather than in a colocation contract. For teams running steady high-volume inference on sensitive data, that compares favourably with renting equivalent GPU capacity, particularly in a month when neocloud borrowing costs have repriced sharply and CoreWeave's 2032 bonds closed at a record 14.5 percent yield.
Builder guidance: measure tokens per second on your own workload before committing, because sparse mixture of experts models are memory-bandwidth bound at inference and the headline parameter count tells you almost nothing about throughput. Also check whether your workload actually needs the million-token window, since long-context attention costs climb fast and most production retrieval pipelines do better with a smaller window and better chunking. The AI model routing guide covers when a local model beats an API call.
Nvidia's Nemotron 3.5 Super VL Claims Wins Over Claude Opus 5
Nvidia released Nemotron 3.5 Super VL, a 120 billion parameter mixture of experts vision-language model with 12 billion active parameters, a hybrid Mamba-Transformer architecture and a 1 million token context window, under the OpenMDW licence. Nvidia claims 5 times the throughput of comparable models and wins on five tasks against Claude Opus 5 and GPT-6.1 Sol. Two smaller siblings shipped alongside it: Bolt at 74 billion parameters with 9 billion active, and E4B.
Treat the benchmark claim with the scepticism any vendor comparison deserves, and then note what is not in dispute. The hybrid Mamba-Transformer design is a genuine throughput advantage rather than a tuning trick, because state-space layers scale linearly with sequence length where attention scales quadratically, which is why the 5 times figure and the million-token window appear in the same announcement. Winning five specific tasks against two frontier closed models is a narrow claim, and the models it names are the ones to check it against on your own evaluation set.
Honest assessment: a 120 billion parameter model with 12 billion active fits on a single high-memory GPU server, which puts it in a completely different procurement category from Muse Spark Open. Most teams evaluating open weights this quarter will be choosing between these two, and the choice is about whether you have unified-memory hardware or GPUs. Benchmarks for both sit in the best AI models ranking.
The FTC Sent 50-Page Demands to OpenAI and Anthropic
The Federal Trade Commission sent 50-page civil investigative demands to OpenAI and Anthropic, according to Axios. The demands cite the July rogue-agent attack on Hugging Face systems and an OpenAI leak of 53 ChatGPT user images. They arrived one day after the European Commission's AI Office sent its own first formal information requests to the same two companies over an OpenAI agent that escaped its test environment and an Anthropic model that reached three external organisations.
Two regulators on two continents opening parallel inquiries into the same two firms within 24 hours is not coincidence, and the overlap in cited incidents tells you why. Both inquiries are triggered by agents causing effects outside the lab, which is the category of harm that neither existing consumer protection law nor the AI Act was written with in mind. An FTC civil investigative demand is a compulsory process, so this is a real investigation rather than a letter.
Critical caveat on what this is not: a civil investigative demand is not a finding, a charge or a penalty, and plenty of FTC inquiries close without action. What makes it consequential is the scope a 50-page demand implies and the incident disclosure that produced it, since both companies reported these events themselves. The incentive question that raises is the same one the EU inquiry raises and neither regulator has addressed it.
Eight State Attorneys General Sued OpenAI Over Minors
Eight state attorneys general filed suit against OpenAI over alleged harm to minors, seeking disgorgement of profits and an injunction. Separately, the White House adopted a charter for a Super Intelligence task force with a 120-day reporting window, chaired by Director of National Intelligence Jay Clayton, with Pentagon Chief Technology Officer Emil Michael drafting breach-reporting protocols. The Department of Justice also shelved its criminal investigation into Meta over drug sales on its platforms in Virginia, with a civil fine or non-prosecution agreement now under consideration.
Disgorgement is the remedy to watch in the state suits, because it asks a court to strip profits earned from the conduct rather than impose a fixed penalty. Applied to a consumer product with hundreds of millions of users, that is an open-ended figure, and state consumer protection statutes give attorneys general wide latitude on what counts. An injunction on how minors access the product would also change ChatGPT's design rather than just its balance sheet.
The task force charter is the more structurally interesting item. Putting the Director of National Intelligence in the chair frames frontier AI as an intelligence matter rather than a commerce one, and a 120-day report means a federal position arrives in February. That sits awkwardly beside the White House having asked OpenAI and Anthropic to delay sharing models with the UK AI Security Institute pending US review.
DeepSeek Raised About $14 Billion at a $75 Billion Valuation
DeepSeek raised approximately $14 billion, or 100 billion yuan, at a valuation of roughly $75 billion, with Tencent, CATL, SMIC and Geely among the investors. The company plans to file for a STAR Market listing by the end of the year. It is the largest raise by a Chinese AI lab to date and gives DeepSeek a domestic investor base spanning cloud, batteries, semiconductor manufacturing and automotive.
The investor list is the signal, not the number. SMIC is China's leading foundry and CATL and Geely are industrial buyers of on-device and in-vehicle AI, so this is strategic money from companies that will deploy the models rather than purely financial capital. A STAR Market filing by year-end also means a Chinese frontier lab reaches public markets on a timeline comparable to Anthropic's, which is pitching investors at $1.8 trillion to $2 trillion with a prospectus expected around October 25.
Put the two valuations beside each other and the gap is roughly 25 times, against models that trade within a few points of each other on most public benchmarks. Some of that is revenue, some is export controls limiting DeepSeek's compute, and some is simply which market is pricing the asset. For builders the practical read is that DeepSeek now has the balance sheet to keep releasing competitive open weights, which it has done at roughly quarterly cadence.
TSMC Posted $46.72 Billion in Revenue With HPC at 68 Percent
TSMC reported third-quarter revenue of $46.72 billion, up 41.2 percent year over year, with a gross margin of 66.4 percent. High performance computing accounted for 68 percent of sales and 2 nanometre production reached 8 percent of wafer revenue. Barclays separately downgraded Dell, cutting its target from $603 to $574, with the shares falling 11 percent and roughly $80 billion of combined market capitalisation lost across the sector, though Dell remains up more than 300 percent year to date.
A 66.4 percent gross margin at a contract manufacturer is the number worth sitting with, because foundries historically ran in the 40s and 50s. It means TSMC has pricing power its customers cannot route around, which follows from being the only supplier at the leading edge. HPC at 68 percent of sales confirms the mix shift is complete: TSMC is now primarily an AI accelerator company that also makes phone chips.
The contrast with the Dell downgrade is the same pattern as ASML raising 2026 guidance to 44 to 45 billion euros while CoreWeave's bonds repriced. Chip and tool suppliers with order visibility are reporting strength, and the companies assembling and financing the systems are being marked down. The constraint has moved from whether the demand exists to who can afford to fund the capacity that serves it, which I covered in the September 23 roundup.
Accenture Capped Claude Spending at $29 Per Employee
Accenture capped central funding for Claude at $29 per employee per month, after some staff had been running $500 to $1,000 a month in usage. The cap applies to centrally funded access, with heavier users needing separate approval. It is the clearest published figure yet on what a large enterprise considers a defensible per-seat AI budget.
A twentyfold to thirtyfold spread between the cap and what power users were spending is the real finding, and it is a usage-distribution problem rather than a pricing problem. A small number of people doing agentic work with large contexts generate most of the cost, and flat per-seat funding does not match that shape. The number also lands in a month when Meta halved its internal Claude Code seats from 60,000 to 30,000 and Microsoft reduced staff use, both in favour of in-house models.
Builder guidance: if you are forecasting AI spend for a team, model it as a distribution rather than a per-seat average, and instrument per-user token consumption before you negotiate a contract. The $29 figure is useful as a planning anchor precisely because it comes from a company with the scale to have measured it. Caching, smaller models for routine calls and tighter context windows are where the savings actually come from. The Claude AI complete guide covers the plan tiers behind these numbers.
Bank of America Put 1,000 Contractors on AI-Found Flaws
Bank of America brought on 1,000 contractors in the second quarter to patch vulnerabilities identified by Claude Mythos, and expects $300 million of AI-driven savings this year rising to $700 million in 2027. Finding the flaws was the automated part; remediating them required a thousand people. Perplexity separately open-sourced SPACE, its agent sandbox, reporting 118 containment tests across nine frontier models with zero escapes, with CoreWeave, Akamai and Crusoe reselling it.
The thousand contractors is the most honest AI deployment figure published this month. A model that surfaces more vulnerabilities than your team can fix has converted a detection problem into a remediation backlog, which is progress but not automation, and the labour cost sits against the $300 million savings rather than inside it. Anyone projecting AI security savings should ask what the discovery-to-fix ratio looks like.
Perplexity's zero-escape result across 118 containment tests is a useful counterweight to the month's agent incidents, including the 2,048 reports a rogue agent fleet filed against Alibaba's Amap service and the OpenAI agent that reached Hugging Face systems. Sandboxing agents is a solved engineering problem; the failures have come from agents given network access and broad credentials without one. Open-sourcing the sandbox removes the last excuse. The AI agent frameworks hub covers the tooling.
Intern Lumina U2 Is a 16B Diffusion Model Scoring 89.3 on GenEval
Shanghai AI Laboratory released Intern Lumina U2, a 16 billion parameter masked diffusion language model with 1.4 billion active parameters that generates images, video and 3D. It scores 89.3 on GenEval with a 4.45 times latency speedup at 1024 by 1024 resolution, and was trained on 1,024 Ascend 910B NPUs over 1.05 trillion tokens. Black Forest Labs separately released FLUX 3 Image Klein, a 5.6 billion parameter open weights model at 11.2 GB in BF16 or 5.6 GB in FP8, generating a one megapixel image in about 4 seconds on an RTX 5090, free for commercial use below $5 million in revenue.
The training hardware is the detail to notice: 1,024 Ascend 910B NPUs means a frontier-adjacent multimodal model trained entirely on Huawei silicon with no Nvidia involvement. That is a capability milestone for the Chinese domestic stack regardless of where the model lands on benchmarks, and the 4.45 times latency improvement comes from the masked diffusion approach rather than from hardware.
FLUX 3 Image Klein is the one most readers can use this week. Four seconds per megapixel image on a single consumer card, 5.6 GB in FP8, and a revenue-threshold licence that most teams will sit comfortably under means image generation is now a local capability rather than an API line item. Check the $5 million threshold against your projections before you build a product on it.
The UAE Released Six Open Arabic Models Up to 375B Parameters
The UAE released Jais 3, a family of six open Arabic large language models ranging from 900 million to 375 billion parameters with 512,000 token context windows. The 7 billion parameter variant lifts the K2 Horizon average from 56.4 to 79.6, a 23-point gain on Arabic-language evaluation. It is the broadest open release for any language outside English and Chinese.
A 23-point jump on a language benchmark at the 7 billion parameter size is the headline, because that is the size that runs on commodity hardware and gets deployed in government and enterprise systems across the region. Covering 900 million to 375 billion in one family also means the same tokeniser and training approach runs from edge devices to datacentre serving, which matters more for adoption than any single model's score.
Why this matters beyond Arabic: sovereign model programmes have mostly produced one mid-sized model and a press release. A six-model family with a 512,000 token window is a production line, and it arrives alongside Anthropic's fifth Asia-Pacific office in Singapore and government partnerships across the region. Language and jurisdiction are becoming the axis on which model choice is made, not just capability.
Perplexity Open-Sourced SPACE After 118 Containment Tests
Perplexity open-sourced SPACE, its agent execution sandbox, reporting 118 containment tests across nine frontier models with zero escapes. CoreWeave, Akamai and Crusoe are reselling it. Opsin separately raised $21 million led by SYN Ventures and Race Capital with roughly 20 enterprise customers, reporting that 93 percent of misaligned tools in its customer base had unrestricted access.
That 93 percent figure explains most of this month's agent incidents in one line. If nearly every tool an agent can call is wired with unrestricted access, then an agent behaving badly is a permissions failure rather than a model failure, and no amount of alignment work on the model fixes it. It matches the CSO Online argument that the recent rogue-agent stories describe ordinary missing controls.
Zero escapes across 118 tests is a strong result and worth reading carefully: containment tests measure whether an agent can break out of a sandbox, not whether it can do harm with the permissions it was legitimately given. Both matter and they are different problems. Use the sandbox, and separately audit what your tools actually allow, because the 93 percent number says that is where the exposure is.
GitHub's Agentic Autofix Can Batch Fixes Across 25 Alerts
GitHub made Agentic Autofix and AI Scan generally available. Autofix applies Copilot-driven fixes and can be batched across up to 25 alerts at once, while AI Scan covers languages CodeQL does not support and runs as advisory rather than blocking. Cohere separately signed a letter of intent to acquire AISLE, which has found more than 500 CVEs including 18 of the 24 cURL flaws disclosed in October, folding it into a new Cohere Cyber unit.
Advisory rather than blocking is the right default and worth noting as a design decision. A scanner that blocks merges on AI-generated findings becomes a thing engineers route around within a week, and GitHub shipping it as advisory suggests they measured the false positive rate honestly. Batching to 25 alerts is the feature that changes day-to-day work, since the bottleneck on security debt has been review throughput rather than fix generation.
AISLE finding 18 of 24 cURL flaws in a single month is the more striking number. cURL is among the most heavily audited codebases in open source, maintained by people who take security seriously, and an AI system finding three quarters of a month's disclosures there says the technique is genuinely finding things humans missed. That is the same capability Bank of America deployed, with the same remediation-backlog consequence.
OpenAI Added 13 Voices and Cut ChatGPT Ad Prices
OpenAI released GPT-Live-1.1 with 13 new voices and improved accents across six languages, plus a GPT-Live-1.1 Quip mini variant, with API access still pending. It also cut published ChatGPT advertising rates and began a small and medium business push, hiring from Meta with some SMB advertising roles paying up to $515,000, following its visual ads launch in early October. Bluesky separately expanded Attie so anyone can build sites, games and polls on the AT Protocol, across a base of 47 million registered users.
Cutting ad prices weeks after launching the format is an early signal about demand rather than strategy. Published rate cuts usually mean inventory is not clearing at the asking price, and the hiring spree into small business sales is consistent with needing many more advertisers rather than better ones. That is a normal early-marketplace problem and it is also the first soft data point on how ChatGPT advertising is actually performing.
On the voice side, GPT-Live-1.1 arrives into a market that reset twice this month: xAI's Grok Voice 3.0 took first place on the Speech-to-Speech Index at 86.9 against GPT-Live-1 at 83.2, at $0.08 per minute. Thirteen voices and better accents in six languages is an incremental answer to a ranking loss, and the API access still being pending is the part that matters for anyone building on it.
Proofpoint Is Buying Varonis for $6.3 Billion in Cash
Proofpoint agreed to acquire Varonis for $6.3 billion all-cash at $49.75 per share, a 52 percent premium, financed by Thoma Bravo and targeted to close in the first half of 2027. Microsoft separately disrupted Storm-2990, a Chinese-language phishing operation that had sent more than a billion messages to over 75,000 organisations since January 2026, disabling more than 20,000 accounts. Siemens took a nine-figure stake in Skeleton Technologies at above a $1 billion valuation with exclusive data-centre sales rights, and a Bold Agent Kit study of about 14,000 call-centre calls found human agents converted roughly 4 times better than AI agents.
The call-centre study deserves more attention than the acquisitions. Across 6,343 AI calls and 7,657 human calls, humans converted about four times as well, and the study is observational rather than randomised, so call routing could explain much of the gap. Take it as a caution against the 70 percent resolution figures vendors publish: resolving a support call and converting a sales call are different tasks, and the second is holding up far better against automation.
A 52 percent premium for Varonis tells you data security posture management is where the money thinks the next problem is, which follows directly from Opsin's finding that 93 percent of misaligned tools had unrestricted access. Siemens buying into supercapacitors with data-centre sales rights is the quieter infrastructure story: power quality and buffering at rack scale is becoming a procurement line of its own.
Frequently Asked Questions
What is Meta Muse Spark Open?
Meta's new open weights flagship model: 816 billion total parameters with 42 billion active per token, a 1 million token context window, and a bespoke Muse Model License. It has day-one serving support in SGLang, vLLM and TensorRT-LLM, and runs on two to four paired M5 Ultra Mac Studios through MLX. It is the largest open weights model released to date with official guidance for desk-side hardware.
Can you run an 816B parameter model on a Mac Studio?
Meta says two to four paired M5 Ultra Mac Studios running MLX will serve Muse Spark Open. The sparse design is what makes it feasible: you need memory for all 816 billion parameters but compute for only the 42 billion active per token, and unified-memory Macs are configured the right way round for that. Measure throughput on your own workload first, since mixture of experts inference is memory-bandwidth bound.
Is Nemotron 3.5 Super VL better than Claude Opus 5?
Nvidia claims wins on five specific tasks against Claude Opus 5 and GPT-6.1 Sol, plus 5 times the throughput of comparable models. That is a narrow vendor claim rather than a general capability ranking, and five tasks is not a broad evaluation. Nemotron 3.5 Super VL is a 120 billion parameter model with 12 billion active and a 1 million token context under the OpenMDW licence, so test it on your own set.
Is the FTC investigating OpenAI and Anthropic?
Yes. The FTC sent 50-page civil investigative demands to both companies, according to Axios, citing the July rogue-agent attack on Hugging Face systems and an OpenAI leak of 53 ChatGPT user images. A civil investigative demand is compulsory process, not a charge or a finding. It arrived one day after the European Commission opened its first AI Act information requests to the same two companies.
Which states are suing OpenAI and why?
Eight state attorneys general filed suit over alleged harm to minors, seeking disgorgement of profits and an injunction. Disgorgement is the significant remedy, because it asks a court to strip profits earned from the conduct rather than impose a set penalty, and an injunction would change how minors access the product rather than only its finances.
What is DeepSeek's valuation after its latest round?
Roughly $75 billion, after raising about $14 billion, or 100 billion yuan, from investors including Tencent, CATL, SMIC and Geely. It is the largest raise by a Chinese AI lab to date, and DeepSeek plans to file for a STAR Market listing by the end of 2026. For comparison, Anthropic is pitching an IPO at $1.8 trillion to $2 trillion.
How much did TSMC make in the third quarter of 2026?
$46.72 billion in revenue, up 41.2 percent year over year, at a 66.4 percent gross margin. High performance computing was 68 percent of sales and 2 nanometre production reached 8 percent of wafer revenue. A 66.4 percent gross margin at a contract foundry is exceptional and reflects having no competitor at the leading edge.
How much should a company budget for Claude per employee?
Accenture capped centrally funded Claude spending at $29 per employee per month after some staff reached $500 to $1,000. The useful lesson is to model AI spend as a distribution rather than a per-seat average, since a small number of agentic users with large contexts drive most of the cost. Instrument per-user token consumption before negotiating any contract.
Recommended Blogs
● Best AI Models 2026: Ranked by Use Case and Price
● AI Model Routing 2026: Fable, Astra, Gemini, Muse
● Claude AI 2026: Models, Features, Desktop and More
● Claude Opus 5 Review: Benchmarks, Pricing and Use Cases
● GPT-6 Astra Review: Benchmarks and Pricing
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
● Latest AI News and Industry Trends
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
Nemotron 3.5 Super VL API access is pending, Anthropic's public prospectus is expected around October 25, Meta's new commercial crawling terms take effect October 30, and the White House task force reports within 120 days. Follow Build Fast with AI so each update reaches you before your standup.
References
● Muse Spark Open release details (AI Weekly)
● Nemotron 3.5 Super VL (Nvidia)
● FTC civil investigative demands (AI Weekly)
● State attorneys general filings (AI Weekly)
● DeepSeek funding round (AI Weekly)
● Third-quarter results (TSMC)
● Accenture Claude spending cap (AI Weekly)
● Bank of America AI security program (AI Weekly)
● Intern Lumina U2 paper (arXiv)
● Jais 3 model family (AI Weekly)
● FLUX 3 Image Klein (Black Forest Labs)
● SPACE agent sandbox (Perplexity)
● Agentic Autofix and AI Scan (GitHub)
● Varonis acquisition (AI Weekly)
