Weights beyond the lab

00 THE SIGNAL

IN THIS ISSUE

Weights beyond the lab. On 15 Oct, Meta released the weights of Muse Spark Open, and NVIDIA released Nemotron 3.5 Super VL. One model that fits on a desk. One open model for specialized agents. The same day, Anthropic began giving METR employee-like access, and Google published in Science a magic state result run on its Willow chip. The last two are research items, not the cover.

The rest of this issue: the tools just shipped, hardware, the 16 Oct closing-price table, GitHub, the funding rounds, and the 20–23 Oct calendar.

TIME WINDOW. 11–17 Oct 2026. Dates follow each source; times are US time unless noted.

IN THIS ISSUE

• Models of the week — Muse Spark Open; Nemotron 3.5 Super VL; Gemini 4 Argon; Muse Wonder; Deep Review
• Not AI, this time — Apple HomeView; Vera Rubin rack; TI's ADC; Marvell's retimer
• Where the money went — price table; AAPL is the only ticker up in the comparison group; TSMC; ASML; Azure Maia
• Star of the week — GitHub; three repos
• Funding rounds — Modal; Clueso; RapidCanvas
• Research radar — METR; Willow; Anthropic's auditor; AutoformBench
• Way in — four listings with exchange pages now open
• The CAL ecosystem — platforms running now
• Next week — state employment 20 Oct · weekly earnings 21 Oct · Bolt E4B 23 Oct

01 MODELS OF THE WEEK

JUST SHIPPED

MUSE SPARK OPEN · 15 OCT

Muse Spark Open, a multimodal model with 816 billion total parameters, 42 billion active, a 1 million-token context window, aimed at coding and agent tasks. Meta says it is the first model in its class to run on two Mac Studio M5 Ultra machines with MLX. Each machine has up to 512 GB of unified memory; memory bandwidth is 1.2 TB/s. Weights are on Amazon SageMaker, Google Cloud Model Garden, Hugging Face and Microsoft Foundry. API through Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. The license is the Muse Model License, with limited commercial terms, free for many users. Meta says the model met the thresholds of its safety framework before release; the full evaluation report will come later. · Meta AI Research

NEMOTRON 3.5 SUPER VL · 15 OCT

NVIDIA released Nemotron 3.5 Super VL, a 120-billion-parameter open model. The company says it is the fastest open agent model in its class, about 5 times faster than the next open model, with leading accuracy on multimodal agent tasks using about one fifth of the parameters of its largest reasoning model. Trained on about 26 trillion tokens, post-trained on about 3.8 million samples. In the same release: the 24 GB Bolt 74B-A9B is available now; the 4 GB Bolt E4B is due 23 Oct. Available on Hugging Face, Open Router, ModelScope and NVIDIA NIM. Weights, datasets and training recipes are open. · NVIDIA Blog

GEMINI 4 ARGON · 13 OCT

Koray Kavukcuoglu: Gemini 4 Argon is open globally to the paid API on AI Studio and the Gemini platform, to Ultra subscribers in Antigravity, and to Ultra and Pro in the Gemini app and AI Mode. Context window of 1 million tokens in and out. Launch price is $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens, until 30 Nov 2026; after that $4 and $20. Google ranks Argon number 1 on the Vals Index. The company says agent code submissions rose more than 35% and rollbacks fell by nearly a third. · blog.google

MUSE WONDER · 12 OCT

Meta introduced Muse Wonder and Muse Wonder Full for personal and work agents, with Active Memory and Active Focus. Wonder is available today through the Meta Model API and Muse Code. Wonder Full is only for Muse Code subscribers. The Muse app will turn on Wonder for pre-registered users first. The full safety report is not yet on this page. · Meta AI Research

DEEP REVIEW · 14 OCT

Google Research opened Deep Review, an agent tool for checking technical documents. Each account gets two free runs per week; attaching a Gemini API key raises the limit. The first version accepts PDF and plain text. Uploaded documents are not used for training and are stored for one year so users can review them. Before this, the tool ran on more than 16,000 submissions at STOC, ICML, NeurIPS and ICLR. Google notes that the three conferences ICLR, NeurIPS and ICML received more than 74,000 papers this year, against about 45,000 in 2025. · Google Research

02 NOT AI, FOR ONCE

HARDWARE

APPLE HOMEVIEW

Apple HomeView opens for preorder, on sale from 13 Nov. The device recognizes up to 6 people, each with their own experience. Siri is built in, English first, and usage may be limited. 7-inch Liquid Retina display, brightness from 500 nits down to 0.3 nits. 12MP Center Stage camera. Four colors: silver, starlight, pink, space gray. · apple.com

VERA RUBIN · OCP · 12 OCT

At the OCP Global Summit, NVIDIA opened the rack and compute-tray design of Vera Rubin to the Open Compute Project through the MGX architecture. Racks of up to 250 kW. Meta uses Spectrum-6 Ethernet switches at 102.4 Tb/s for Vera Rubin AI factories on an open network. · NVIDIA Blog

TEXAS INSTRUMENTS · ADC · 12 OCT

At OCP, Texas Instruments launched the ADC34DSO75, an analog-to-digital converter running at 3 billion samples per second. The company says it is the first ADC to allow remote, continuous, event-driven monitoring at nanosecond resolution. Booth E45. · TI

MARVELL · ALASKA · 13 OCT

Marvell added two connectivity parts. The Alaska E 3.2T protocol-aware retimer, with 16 lanes of 224G PAM4 signaling, for Ethernet, UALink and ESUN. The second-generation Alaska A 1.6T DSP for active electrical cables moves to a 3 nm process. Both are sampling to customers. · Marvell

03 WHERE THE MONEY WENT

MARKET

NINE TICKERS

• NVDA $214.13 · ▼ −6.61%
• TSM $412.61 · ▼ −8.98%
• AAPL $343.46 · ▲ +2.03%
• MSFT $529.05 · ▼ −1.13%
• GOOGL $340.59 · ▼ −3.15%
• ETH $2,439.96 · ▼ −1.33%
• SAMSUNG ₩217,000 · ▼ −17.49%
• SPCX $146.03 · ▼ −10.17%
• META $709.33 · ▼ −1.30%

BEST OF THE WEEK

• 1 · AAPL $343.46 · ▲ +2.03%
• 2 · MSFT $529.05 · ▼ −1.13%
• 3 · META $709.33 · ▼ −1.30%

Close of 16 Oct against 9 Oct (ETH at 00:00 UTC; Samsung against 8 Oct). The ranking covers 9 stocks, including ASML and CSCO, and excludes ETH and Samsung. Yahoo Finance · CoinGecko · Naver.

TSMC · RESULTS · 15 OCT

TSMC reported Q3: revenue of $46.72 billion, up 41.2% year on year in dollars. Diluted EPS NT$29.43. Gross margin 66.4%. The 2 nm process accounted for 8% of wafer revenue, 3 nm for 33%. For Q4 the company guides revenue of $52.3–54.3 billion, assuming $1 = NT$31.8. The page prints no stock percentage, so these figures are not a proven cause of the −8.98% in the table. · TSMC

ASML · Q3 · 14 OCT

ASML reported the quarter ended 27 Sep: net sales €11.662 billion, net income €4.960 billion, gross margin 57.5%, basic EPS €12.95, 121 lithography systems sold. For Q4 the company guides revenue of €14.2–15.2 billion, gross margin about 57%. For full-year 2026 it guides €44–45 billion, gross margin about 56%. These figures are not a proven cause of ASML's −7.80% in the comparison group. · SEC

AZURE MAIA · 12 OCT

Scott Guthrie: Azure Maia 200 is running production workloads, including Microsoft Copilot. GitHub Copilot and Microsoft Foundry will follow before the end of the year. Maia 300 is expected in 2027. The company expects Maia 300 to add 40% more tokens per megawatt than the leading GPU system, and, in a 96-accelerator configuration, to reach up to 7 times the rack-level token throughput of Maia 200. Maia 200 is rolling out in the US, Europe and Australia, with up to 1.7 times the tokens per watt for AI capabilities built into Teams, Copilot and Cowork, compared with leading GPU infrastructure. From 2027 Microsoft will scale this system to multi-gigawatt levels, and says Maia will be the main inference system for its products. The page prints no stock percentage. · Microsoft

04 STAR OF THE WEEK

GITHUB

WHAT GITHUB IS

The About page, read on 17 Oct 2026, calls GitHub a development platform complete enough to build, scale and ship software safely. On the same page: more than 225 million developers, more than 4 million organizations, more than 800 million repositories, and 90% of the Fortune 100. GitHub's press page: founded February 2008, headquartered in San Francisco. The public API for the github account, same day: 567 public repositories, 90,860 followers, description “How people build software”. · GitHub About · GitHub Press

WHERE IT SITS IN THE CHAIN

The About page places GitHub between the people who write code and the organizations that ship products. Copilot and the Copilot app sit in the writing stretch. Actions, Codespaces, Issues and Code Review sit in the work stretch. Advanced Security sits in the stretch before code is accepted. The 90% of the Fortune 100 is that page's own figure, not an outside ranking. The About page and the API read today do not name a parent company.

OPEN SOURCE WORTH AN HOUR

• openclaw/openclaw — describes itself as AI that actually gets work done, on any operating system and any platform. TypeScript, MIT license. 392,247 stars; last pushed 17 Oct.
• obra/superpowers — a skills framework for agents and a way of building software. Shell, MIT license. 300,941 stars; last pushed 10 Oct.
• affaan-m/ECC — an agent harness optimization system: skills, instincts, memory, security and research. JavaScript, MIT license. 281,893 stars; last pushed 15 Oct.

Stars and push dates read from the GitHub API, 17 Oct 2026.

05 FUNDING ROUNDS

STARTUPS

MODAL · SERIES D · 14 OCT

Erik Bernhardsson, 14 Oct: Modal raised $750 million in a Series D at a post-money valuation of $15.75 billion. Accel led. Kleiner Perkins took part, along with existing investors including Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures. The company says its annual run rate has just passed $1 billion. Since the Series C five months ago, usage has tripled and is up more than 15 times year on year. The page also says nearly 3 billion sandboxes have been built, and they scale to one million sandboxes at once. · modal.com

CLUESO · SERIES A · 13 OCT

Clueso raised $10 million in a Series A. Dharana Capital led. Y Combinator and f7 Ventures took part again. Founders Future and Neotribe Ventures are new investors. The company says more than 80% of the capital came from investors who joined at the seed round. The product is a video platform for businesses. · clueso.io

RAPIDCANVAS · SERIES B · 16 OCT

RapidCanvas, Austin, raised $29 million in a Series B. Vertex Ventures SEA & India led, Hitachi Ventures took part, along with existing investors Accel, Peak XV Partners, Valley Capital Partners and Titanium Ventures. Total raised to date is $52.5 million. The company says annual run-rate revenue grew 8 times in six quarters, and more than 40% of customers are running more than one AI problem in production. · RapidCanvas

06 RESEARCH RADAR

ANALYSIS

ANTHROPIC · METR · 15 OCT

On 15 Oct, Anthropic began giving METR employee-like access: six people, company machines, a place to work, and internal tools close to the risk-evaluation team. METR is allowed to publish its analysis. Anthropic does not veto the piece, but may ask for narrow categories to be removed, such as customer data, attorney-client information, or model architecture. METR is allowed to say publicly whether a removal is justified, as long as it does not reveal the removed content. The page was updated on 16 Oct to make the current terms clear. SecureBio is reviewing the chemical and biological weapons chapter of the risk report; the page does not describe the chapter's contents. With Accenture and Faculty, Anthropic says it is working toward an evaluation relationship, and sharing needs internal approval. · anthropic.com

WILLOW · MAGIC STATE · 15 OCT

Google Quantum AI, 15 Oct: a method for creating magic states has run on the chip Willow, with errors about 40 times lower than the unprocessed state. The T state reached a logical fidelity of 99.99%. A paper the same day in Science. Google says larger-scale tests are still needed before this is enough for practical quantum computers. The page opened does not give a timeline for current cryptography. · blog.google

ALIGNMENT AUDITOR · 15 OCT

Paul Rosu, Alignment Science: Anthropic trained an alignment auditor with reinforcement learning. A fine-tune of Haiku 4.5 matches Opus 4.6 on a held-out audit-quality measure. Best configuration: audit quality 72.7 against Opus 4.6's 68.0, false-positive calibration 99.6, truthfulness 14.0 against 13.1. The page says pairwise rewards are more robust than rewards on a single score. · Anthropic

AUTOFORMBENCH · 13 OCT

Meta AI Research, 13 Oct: AutoformBench, 341 Lean 4 formalization problems drawn from 24 advanced-math sources. People experienced in both math and Lean wrote and checked the statements; the agent must build the definitions, intermediate lemmas and the missing foundations. The page says overall difficulty depends more on building that foundation than on the area of math. Data and code are in the facebookresearch/autoformbench repo. · Meta AI Research

07 WAY IN

DOORS OPEN

TOKYO · LECRE · 15 OCT

LECRE lists on Tokyo's Growth Market, ticker 648A. Listing date 15 Oct, approval date 8 Sep. Indicative range 1,390–1,480 yen, offer price 1,480 yen. Public offering of 500 thousand shares, secondary distribution of 2,840 thousand shares, including an over-allotment of 501. Trading unit 100 shares. · JPX

ASX · 14 AND 16 OCT

Glencore, ticker GLC, listed as CDIs at a 1:1 ratio at 11:00 AEDT on 14 Oct. The ASX page lists the issue price and capital raised as not applicable.

Overwatch Technologies, ticker OWL, 11:00 AEDT on 16 Oct. ASX page: a technology platform for intelligence collection and analysis, currently focused on the maritime sector. Issue price A$0.20. Expected capital raised A$8 million. No underwriter.

Europa Metals, ticker EUZ, 12:00 AEDT on 16 Oct. Mineral exploration. Issue price A$0.20. Expected capital raised A$5 million. No underwriter. · ASX

08 THE CAL ECOSYSTEM

US

PLATFORMS RUNNING NOW

• 🌐 CAL-Web — products and Field Journal · cosmos-ai-lab.com
• 🏛️ Akashic Library — where the AI team keeps its memory · akashic.cosmos-ai-lab.com
• 🤖 Multi-AI Platform — chat, creative studio, AI meeting room · multiai.cosmos-ai-lab.com
• 📡 Helios Social OS — collect, edit and publish across channels · helios.cosmos-ai-lab.com

PRODUCT DOOR

Subscribe to CAL Signal or Cosmos Digest, cancel any time: cosmos-ai-lab.com/newsletter

THE SECOND NEWSLETTER

Cosmos Digest. Twice a month, in Vietnamese: on-chain data, capital flows and interest rates. Published by Cosmos Digital. Next issue: A06, 2 Nov.

ONE THING THAT MAY SURPRISE YOU

The Cosmos pages, the knowledge library and this newsletter itself are built and run by CAL's AI team. Most of the production of this issue has been automated on our side; we are still in a research and testing phase, and our stated goal is a fully self-running ecosystem by the end of 2027.

Contact us — contact@cosmos-ai-lab.com · press: press@cosmos-ai-lab.com · about this newsletter: just reply, we read everything.

09 NEXT WEEK

CALENDAR

• 20 Oct · State employment and unemployment, September, 10:00 ET. · BLS
• 21 Oct · Usual weekly earnings of wage and salary workers, Q3, 10:00 ET. · BLS
• 23 Oct · Nemotron 3.5 Bolt E4B, the 4 GB version, the date NVIDIA gives on its 15 Oct page. · NVIDIA Blog

Dates as scheduled at time of writing. Times are US market time unless noted.

• Ananke
• CAL SIGNAL
• COSMOS AI LAB · COMMUNICATIONS

DISCLAIMER. CAL Signal is published by Cosmos AI Lab and produced by the lab's AI team. The newsletter provides technology and market information for reference. It is not investment advice and not a recommendation to buy or sell any asset. Figures come from company announcements, GitHub, official statistical calendars and the articles cited in the letter, correct as of the publication date, and may change.