Meta Opens Muse Spark, Argon Hits Antigravity Pro & Quantum Cultivates Magic States
Dispatches of the past four-and-twenty hours from the laboratories of artificial intelligence
Open Weights
Meta Opens the Weights on Muse Spark
816B total / 42B active, 1M-token context—and Artificial Analysis crowns it the top U.S. open-weights model
By Our Menlo Park Desk
Artificial Analysis puts Muse Spark Open (xhigh) at 39 on the Intelligence Index—#1 among U.S. open-weights models.Artificial Analysis שמה את Muse Spark Open (xhigh) על 39 במדד ה־Intelligence—מקום 1 בין מודלים פתוחים בארה״ב.Image: תמונה: @ArtificialAnlys on X
Mark Zuckerberg posted the news in one line: Muse Spark Open—weights releasing today—runs great on new Mac Studios thanks to work with Apple. Meta’s AI account filled in the card. Muse Spark Open is billed as the most capable U.S. open-weight model to date, a mixture-of-experts built for coding and agentic work: long-horizon tasks, hard software engineering, and multi-step reasoning.
The specs are public: 816 billion total parameters with 42 billion active, and a 1-million-token context window. Meta says it is optimized for two Apple M5 Ultra Mac Studios on the latest MLX stack, and also runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp and Unsloth. API access landed at Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI; weights sit on SageMaker, Google Cloud Model Garden, Hugging Face and Microsoft Foundry. Artificial Analysis scored Muse Spark Open (xhigh) 39 on its Intelligence Index—#1 among U.S. open-weights models, 22 points above Meta’s own Muse Glimmer. Ollama shipped cloud access the same day (ollama run muse-spark-open:cloud), and Cursor already hosts Muse Wonder from Meta Superintelligence Labs.
A separate weekly Argon quota—so Flash stays free for the rest of the week’s work
From the Antigravity Beat
Google’s Antigravity team is rolling Gemini 4 Argon out to paid Google AI Pro subscribers today. Demand forced deliberate capacity scaling; Pro now gets a weekly-only Argon quota on top of the existing Gemini bucket for 3.8 Flash. Demis Hassabis and Sundar Pichai amplified the news. ARC Prize’s verified numbers put Argon at 97.5% on ARC-AGI-3 with a provider adapter harness—and François Chollet noted its ARC-AGI-3 profile sits near Astra. Docs: antigravity.google/docs/models.
New language voices across ChatGPT, Codex and dots—plus computer use for anyone with the Chrome extension
By Our San Francisco Desk
OpenAI shipped GPT-Live-1.1 to Voice in ChatGPT, Codex and dots: new voices in French, German, Hindi, Mexican Spanish, Korean and Japanese, better accents, and less of the dreaded “checking…”. Rolling out to all users today; API soon. Separately, ChatGPT brought computer use to Chrome users. In Chat it can read across open tabs; in Work it can drive tabs—send messages, fill forms, download files, even group the tab pile. The in-tab side panel now carries full ChatGPT history. Extension is free for all plans.
First hardware demo of magic-state cultivation—40× fewer errors, 99.99% logical fidelity, in Science today
Science from Mountain View
Sundar Pichai flagged “magical progress” toward useful error-corrected quantum computers. Google Quantum AI demonstrated magic-state cultivation on physical hardware for the first time on the Willow processor: a 40× error reduction versus prior methods and state-of-the-art 99.99% logical fidelity. The result is published in Science; Demis Hassabis boosted the thread.
Opus 5.5 Fast Mode Exits Preview; Code Review Goes GA
Same Opus, up to 2.5× faster at $8/$40 per million—and Team/Enterprise get Quick, Standard or Deep PR reviews
From the Claude Desk
ClaudeDevs pulled Opus 5.5 fast mode out of research preview and killed the waitlist. Same Opus 5.5 quality, up to 2.5× faster, priced at $8 / $40 per million tokens—also on Microsoft Foundry (Azure), with Google Cloud soon. Hours later, Code Review hit general availability for Claude Team and Enterprise: ~3× faster than preview, tighter findings (problem, trigger, fix), and admin-picked Quick/Standard/Deep levels per repo. Enterprise orgs can claim at least $10,000 in Code Review credits by January 15.
Code Review GA: Quick, Standard or Deep per repo—and findings that name the bug, the trigger and the fix.ביקורת קוד ב־GA: Quick, Standard או Deep לפי ריפו—וממצאים שמציינים באג, טריגר ותיקון.Image: תמונה: @ClaudeDevs on XCodex Ops
Day 10 Reset, Day 11 Capacity—Codex Keeps Shipping
Tibo’s reset button returns; capacity errors should self-heal—and the Codex app gains composer predictions
OpenAI Engineering Notes
thsottiaux (Tibo) spent Days 10 and 11 of his public Codex diary putting fires out. Day 10: “The reports of my reset button being lost are greatly exaggerated”—and it landed within two hours. Day 11: more compute and changes to slash the “Selected model is at capacity” blocker; when it still hits, the run should eventually continue itself. Codex Changelog also shipped composer follow-up predictions (Tab to accept as draft, beta for ChatGPT Pro on GPT-6 Astra and 6.1 Sol) plus a pile of app and CLI polish.
Simons Puts $15M on AI-for-Math; Anthropic Opens the Lab Door
OpenAI and Anthropic each pledge $5M; METR gets employee-like access; two new Anthropic fellowships open
Policy & Philanthropy
The Simons Foundation announced a $15 million push—workshops, initiatives, and free AI compute for foundation-backed researchers—with OpenAI and Anthropic each committing $5 million plus model access. Separately, Anthropic said it has given employee-like access to a METR evals team while talking to other third-party safety assessors. Applications also opened for STEM Fellows and Conceptual Reasoning Fellows: four months at Anthropic from January 2027 to help Claude reason better across science, engineering, philosophy, cognitive science, economics, policy and law.
One model on billions of simulated collisions—for reconstruction, ID, anomalies and future detectors
European Lab Wire
Mistral AI and CERN are exploring what frontier-scale models can do for particle physics. The ambition: a single Large Physics Model trained on billions of simulated collisions, aimed at reconstructing events, identifying particles, flagging unexpected signals, and evaluating future detector designs.
TIME Inventions, Rosalind Labs, Gemma Docs & Hermes
SynthID and Omnilingual ASR honored; biology access expands; NVIDIA models land in Hermes
Roundup from the Wire
Google DeepMind’s SynthID—watermarking for AI media, already on 180B+ images and videos—joined TIME’s Best Inventions of 2026; Meta’s Omnilingual ASR (1,600+ languages) got the same nod. OpenAI opened a Rosalind Academic Plan: up to 12 months of sponsored GPT-Rosalind access for eligible U.S. university labs, with research data not used for training by default. googlegemma shipped a rebuilt docs home with a global model selector and better agent snippets. Nous Research says Hermes Agent is the sweet spot for NVIDIA’s new Nemotron 3.5 Super VL and Bolt models (Super VL hit 63.44 on Hermes Bench). Simon Willison flagged a striking claim: if distillation is 5–7% of Anthropic revenue, that is hundreds of millions other labs spend copying Claude.
Wren & Robin read you this morning's dispatches, with banter.
Edited & published by Nadav Solomon
בשידור · רדיו
האלחוט של הגוזל
רֶן ורובין מגישים את ידיעות הבוקר, עם קצת צחוקים.
בעריכת והוצאת נדב סולומון
Published by Nadav Solomon · Printed on the Fledgling Press · Set in Georgia & Times
By Post · Weeklyבדואר · שבועי
The Fledgling Weeklyהגוזל השבועי
The week's best AI stories and Repo Corner picks, once a week.מיטב ידיעות ה-AI של השבוע ובחירות פינת הריפו, פעם בשבוע.
Newsletter language:שפת הניוזלטר:
We'll email you a confirmation link first. Your address is used only for this newsletter, and your approximate country (from your IP) is stored to pick your delivery day. Never shared; every issue has a one-click unsubscribe.קודם נשלח לך קישור לאישור. הכתובת משמשת רק לניוזלטר הזה, ונשמרת גם מדינה משוערת (לפי כתובת ה-IP) לבחירת יום המשלוח. לא מועבר לאף אחד, ובכל גיליון יש הסרה בלחיצה אחת.
The site uses Google Analytics to count visits; its cookies are set only after you accept in the cookie banner. Privacy policy.האתר משתמש ב-Google Analytics לספירת ביקורים; העוגיות שלו נשמרות רק אחרי אישור בבאנר העוגיות. מדיניות פרטיות.
