Meta Releases Muse Spark Open, a 816B Open-Weight Model, as Perplexity and OpenAI Ship Agent Updates

Key Takeaways

1. Meta released Muse Spark Open, an 816B-parameter open-weight MoE model that NVIDIA says leads US open-weight models on the Artificial Analysis Intelligence Index at 39.
2. Perplexity open-sourced its pplx-decider-v1.2 decision models and added a Plaid connector plus a reverse-engineering skill to Computer.
3. OpenAI upgraded GPT-6 Astra Pro in ChatGPT with parallel agents and progress tracking for Pro users.
4. Anthropic said Claude Mythos 5 can design quantum circuits from natural-language instructions, beating published best circuits in some cases.
5. Musk detailed Grok Bot speed gains and X DM integration, while X set a new advertising aesthetic standard.
6. A researcher-built AdaJEPA adapts world models during planning, and a new survey maps roughly $3B in AI existential-risk funding.

1. Open-Weight Models and Agent Platforms

1. NVIDIA congratulated Alexandr Wang and the Muse team on Muse Spark Open, an open-weight MoE model with 816B total and 42B active parameters that scored 39 on the Artificial Analysis Intelligence Index, which NVIDIA called first among US open-weight models; NVIDIA said Meta provided benchmark access before public release. The score is 22 points above Meta's August 30B dense Muse Glimmer (high, 17) and above NVIDIA's same-day Nemotron 3.5 Super VL (26), while landing within 1 point of Meta's flagship Muse Spark 1.2 (xhigh, 40); it is stronger on AutomationBench-AA with a lower hallucination rate but weaker on GDPval-AA v2.1, CritPt and AA-Omniscience accuracy. 1
2. On efficiency, Muse Spark Open uses about 40k output tokens per Intelligence Index task, the lowest among the nine highest-scoring open-weight models — 16% less than Muse Spark 1.2 (48k) and 43% less than GLM-5.3 (max, 71k); at $1.25/$4.25 per million input/output tokens (cached input $0.15) on platforms such as Together AI, that is $0.84 per task, below Muse Spark 1.2's $0.97 and GLM-5.3's $2.01, but about three times DeepSeek V4.1 Flash's $0.27. It ranks last among the nine on all four agentic evaluations and fourth on the other six combined, while leading all open-weight models on AA-Omniscience at 25; GLM-5.3 (max) leads by 6 points entirely from agentic evaluations. 1
3. Deployment details: BF16 weights are about 1.6TB, while Meta's NVFP4 checkpoint quantizes only 256 experts down to about 530GB and is said to run on four Blackwell Ultra GPUs or eight H200s, or on two 512GB M5 Ultra Mac Studios via MLX; the architecture has 62 layers, grouped-query attention with 16 KV heads, 8 of 256 experts active per token, and a 1.8B-parameter perception encoder, with a 1M-token context, text/image/video/audio input and text output. The custom Muse Model License requires a "Built with Muse Spark Open" label, and model-as-a-service businesses with annual revenue above $20M must sign a separate Meta agreement; BF16 and NVFP4 weights are on Hugging Face, with APIs via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. 1
4. Perplexity open-sourced pplx-decider-v1.2 (27B, 2 cents per million tokens, second overall on the HF Decision Index) and pplx-decider-flash-v1.2 (4B, 1 cent per million tokens, first in the 4B class), citing higher text and vision scores, inference up to 5x faster, and all questions in a single request passing at once; both are on the Perplexity Decisions API, Hugging Face open weights, OpenRouter and Vercel AI Gateway, with OpenRouter called the best platform for the Decider models. Perplexity also added a Plaid connector to Computer for use as a "personal CFO" and made REA (Reverse Engineer Anything) a Computer skill that analyzes an app without source code and builds a working version, with /reverse-engineer tasks free until October 31 up to 10,000 credits per user. 1 2 3 4 5 6
5. OpenAI upgraded GPT-6 Astra Pro in ChatGPT with faster responses, improved plugins and tool calls, and multiple agents working in parallel on parts of a task, with users able to track each agent's progress and open completed tasks for results; it is available to Pro users in the Chat tab. OpenAI also relayed that engineer Ali used dot at 3 a.m. to triage an on-call incident, with dot investigating and creating a Codex thread so Ali could direct the work from the ChatGPT app on his phone. 1 2

2. Research, Science and Developer Tooling

1. Anthropic's STEM Fellows Program said Claude Mythos 5 can design quantum circuits from natural-language instructions, outperforming the best published circuits and in some cases improving by an order of magnitude or more; the account said more efficient circuits can reduce the resources quantum algorithms need, potentially bringing chemistry and materials-science applications of quantum computing closer. 1
2. A relayed research description said world models can fail when the world changes, so Ying Wang, Yann LeCun and Mengye Ren built AdaJEPA to adapt during planning; another relayed item said a new survey mapped nearly $3B in historical and projected funding behind the AI existential-risk ecosystem. 1 2
3. Ethan Mollick recommended a paper on a Gemini 4 agent team completing Minecraft, calling it a readable case of agents spontaneously coordinating toward a complex goal and noting its methods differ from humans'; he also questioned whether current algorithms can filter out obviously AI-generated junk replies. 1 2
4. Deedy said existing large models are either insufficient or too costly for scaled visual-reasoning tasks, and that the Elorian team — described as including people who worked under Jeff Dean on Gemini, Apple's LLM lead, xAI's lead and a Harvard professor — released Glimpse, a visual-reasoning model claimed to reach Astra-level performance at 17x lower cost, which he called important for video-input uses such as manufacturing, robotics, games and content moderation; a quoted post described Glimpse as the team's first "visual thinking" model with frontier performance on key public visual-reasoning benchmarks, aimed at models that handle single images well but fail under motion or viewpoint changes. 1
5. NVIDIA praised Perplexity's open-source agent sandbox SPACE, in which each agent task runs in its own Firecracker microVM with credentials and outbound traffic controlled outside the VM, released under Apache 2.0, saying it helps build safer autonomous AI agents; NVIDIA also thanked partners for post-training customization of Nemotron 3.5 Super VL and relayed that the model is live on Perplexity. 1 2 3

3. Consumer AI, Agents and Platform Moves

1. Musk said Grok Bot now connects to X direct messages and message requests to manage the inbox, which he has used to review new requests for a week and called the best use; he reported computer-task completion time cut to one third and other replies to under half, a simpler chat-card layout, and work on a business-focused Hagglebot targeting more than 5% enterprise cost-efficiency gains and up to $6T a year in savings at global GDP scale. A user said Hagglebot served two large clients and found hundreds of "guaranteed delivery" overnight delays, expecting tens of thousands of dollars in savings, while a relayed account said SpaceXSI's internal Haggle Bot identified over $100,000 in savings in its first week across supplier contracts, idle software subscriptions and recurring purchases. 1 2 3 4 5 6 7
2. X will require all ads to meet "reasonable aesthetic standards," barring ads that are ugly, jarring, shrill or exaggerate product claims; Musk backed the rule and urged users to report ugly or misleading ads to @Ads. Separately, Starlink said it has worked with Indian regulators, industry and operators for over five years, built dedicated systems and gateways in India and obtained necessary government approvals, and is ready to provide high-speed internet there; a relayed account said Reliance welcomed Starlink competition, that Starlink has over 11,000 LEO satellites and 20 Indian gateway sites with a data-stays-in-India commitment, and that Jio has not launched its own LEO satellites, with a first test launch reportedly targeted for September 2027. 1 2 3 4
3. A relayed description of the new Tesla Roadster said SpaceX cold-gas thrusters could enable 0–60 mph in 0.55 seconds, with a 200L carbon-fiber-wrapped pressure vessel behind the seats storing 10,000 PSI nitrogen, thrusters producing 80 kN — comparable to a single fighter-jet engine — and current fill time of 6–7 hours that Tesla aims to cut to 30 minutes, plus Optimus leg actuators managing rear-wing downforce; another relayed item said the car could autonomously perform any demonstrated maneuver and that a "crazier than Hot Wheels" loop test track would be built at Texas headquarters. 1 2
4. Runway launched Mirelo SFX 2.0, which generates a full soundscape synchronized to every video frame while preserving existing dialogue. Demis Hassabis relayed that his models and tools work with researchers worldwide on earlier disease detection, treatment and prevention, including locating rare genetic variants and accelerating drug discovery, and said Antigravity quotas for Argon and Gemini would be reset for everyone after heavy interest this week. 1 2 3
5. Ben Tossell relayed that WebGPU support is coming to uzu with an early preview, that Delegations is generally available so teams can hand work from existing tools to Droid, and that an open-source Claude Code project fixed copy-paste, links and other interactions; he also described spending a full day building his agent setup, with @bot opening threads in @get_bb_app for formal coding work, emailing bot updates, and updating a shared task list and work log, built on @RhysSullivan's Executor with domain, expiry, host and linked local repositories. 1 2 3 4 5

4. Disputed and Unverified Claims

1. disputed/unverified — Yann LeCun disputed several Trump claims with cited figures: oil at $4.35 versus a Biden-era average of $3.46 and $3.01 at departure; factory construction at a record under Biden and down 31% since Trump took office; business investment of $14.5T under Biden versus $6.7T under Trump so far, against Trump's "$21T" claim; and prescription drug prices down only 3.9%, not 300–700%, with US prices still near the world's highest. Relayed opinion also disputed Trump's response to Ukrainian strikes on Russian refineries and alleged Russian election interference, including GRU disinformation actions and expected fake-news campaigns against pro-European candidates; these are source claims, not independently verified here. 1 2 3 4 5 6 7
2. disputed/unverified — LeCun responded to Musk's compressed-air vehicle concept by estimating that if air were 50% of vehicle weight, effective thrust would last under 10 seconds, likely about 5, requiring a hollow, lightweight shell with a large compressed-air tank; he called the demo "cute but completely impractical" and said Musk knows it. Separately, swyx described his team publishing about 2–10 real technical videos a day, launched a "Compute Trading" track with @affaan and @absurdistphil, citing Jump and DRW putting traders and capital into emerging compute markets, and launched "HIGH TASTE TESTERS" comparing tools that work with those that break, covering OpenAI DevDay, Dots/Muse/GrokBot/Instinct, coding and personal assistants, the gap between model capability and usable products, subagents, context management, prompt injection and assistants that can spend money. 1 2 3 4 5 6