NVIDIA Ships Nemotron 3.5 Super VL as Grok Voice 3.0 Tops Speech Index

Key Takeaways

1. NVIDIA released the open-weight Nemotron 3.5 Super VL (120B total / 12B active) for long-running agents, with third-party testing showing it leading US open-weight models and cutting output tokens by roughly 60%.
2. xAI's Grok Voice 3.0 debuted at the top of the Artificial Analysis Speech to Speech Index at 86.9, and SpaceX says Starlink call centers already handle 20,000 daily calls with about 70% needing no human intervention.
3. OpenAI said more than half of EU enterprise and business AI usage now involves agents, with small-business users up 8x, and committed $5M to a Simons Foundation program for mathematicians to review AI-generated results.
4. Anthropic opened applications for two new Fellowships — STEM and Conceptual Reasoning — with selected fellows starting January 2027 for four months, applications due October 28.
5. Perplexity Computer added Ultrafast GPT-6 Astra and SOL 6.1 for Max subscribers and integrated the Polymarket API for prediction-market data.
6. Hugging Face highlighted open-weight releases including Liquid AI's LFM2.5-24B-A2B, which tops the IFEval leaderboard at 96.75, and FLUX 3 Image Klein, a 5.6B model runnable on a single consumer GPU.
7. Modal crossed $1B annualized revenue and raised a $750M Series D led by Accel at a $15.75B valuation, with usage up more than 15x year over year.

1. AI Models and Infrastructure

1. NVIDIA released Nemotron 3.5 Super VL, an open model combining reasoning and tool use with text, image, document, and video understanding for long-running AI agents, at 120B total parameters and 12B active. Third-party evaluation put it at 26 (or 26.5) on the Artificial Analysis Intelligence Index, making it the leading US open-weight model, roughly double the prior Nemotron 3 Super's 13, with about 60% fewer output tokens per task and an NVFP4 version within 0.1 points of BF16. NVIDIA also shipped Nemotron 3.5 Bolt 74B-A9B 24GB, the first Bolt-series model, scoring 21 on the Intelligence Index — 1.5 points below Nemotron 3 Ultra 550B A55B's 23 — with a 256k context window and vision-plus-text input. NVIDIA said it has more models coming this week. — via 1 2 3 4
2. NVIDIA congratulated StepFun on Step 5 Preview, calling it the company's largest open-weight release: a 600B MoE model with 27B active parameters per token, a 1M-token context window, native text and vision input, and targeting agentic work in coding, software engineering, professional knowledge, and finance for tasks up to 24 hours. StepFun said BF16 and FP8 weights are on Hugging Face and ModelScope, and NVIDIA reported it runs at nearly 8K tokens/sec/GPU on GB300 with NVFP4. — via 1
3. Hugging Face highlighted two open-weight releases: Liquid AI's LFM2.5-24B-A2B, a 24B-total / 2.45B-active model scoring 96.75 on IFEval to top the Papers with Code instruction-following leaderboard ahead of Rufus-Air (95.4) and Qwen3.5-27B (95); and FLUX 3 Image Klein, a 5.6B-parameter open-weight image model runnable on a single consumer GPU with native 4K, bounding-box, and pixel-precise editing. — via 1 2 3
4. Modal crossed $1B in annualized revenue and raised a $750M Series D led by Accel at a $15.75B valuation, with usage tripling since the C round and growing more than 15x year over year. One investor said Modal grew 10x from nine-figure revenue in under a year, and a forwarded post noted Modal's first "real" user in early 2022 ran large-scale weather data for about $500, while this year Modal scaled to millions of concurrent sandboxes and built a new sandbox runtime. — via 1 2 3

2. Voice, Agents, and Enterprise Adoption

1. Grok Voice 3.0 launched at the top of the Artificial Analysis Speech to Speech Index with 86.9, ahead of GPT-Live-1 (Astra, medium) at 83.2 and Gemini 3.8 Live Extended Thinking at 82.6, and up 5.6 points from Grok Voice 2.0's 81.3. It ranked 3rd in Speech Agent Arena preference (Elo 1070) with a leading 94.8% task success rate, 71.6% on Tau Voice, 98.5% on Big Bench Audio reasoning, and a first-audio average latency of 0.81 seconds — slower than Grok Voice 2.0's 0.70 seconds. Pricing is $4.80 per hour of input audio, unchanged from Grok Voice 2.0, above GPT-Live-1 (Sol, low) at $4.47 and Gemini 3.8 Live at $3.50, and about 2.2x cheaper than GPT-Realtime-2.1 High at $10.75. SpaceX said Starlink call centers use Grok Voice for 20,000 customer calls per day, with about 70% needing no human intervention and a quarter of sales calls completing a purchase before hang-up; Grok Voice 3.0 also handles millions of daily conversations in Tesla vehicles. — via 1 2 3 4 5 6 7
2. Grok Bot expanded to detect phishing emails with hidden links and generate reports, support email aliases, and give each account a dedicated @mail.grokbot.com address; it can review legal documents such as contracts, autonomously book dentist appointments, and book flights. Musk said SpaceX will fully reimburse users if Grok Bot shopping errors or leaks card information, and after Amazon blocked Grok Bot orders he suggested having the bot comparison-shop elsewhere, noting his own bot bought directly from Lego's website. Users reported connecting email, calendar, and custom apps made it "just work," with one saying a task expected to take about ten days was done in under two hours on the $30/month premium plan, another saying their bot built a fully autonomous business that earned $1,100, and another selling a first product within six minutes of posting a video. — via 1 2 3 4 5 6 7 8 9 10 11 12 13
3. OpenAI said more than half of AI usage among its EU enterprise and business customers now involves agentic tools, with small-business users in the region growing 8x, and its latest economic research report examines Europe's shift to agentic AI and how to turn broad adoption into deep usage. OpenAI also committed $5M to a new Simons Foundation program supporting mathematicians in reviewing, understanding, and building on AI-generated results, including recent OpenAI outputs; the funds are managed and disbursed by the Simons Foundation for workshops, conferences, and special projects, and OpenAI will provide models and compute to foundation-supported researchers. A cited post said the Simons Foundation committed $15M through separate partnerships with Anthropic and OpenAI for workshops, initiatives, and free AI compute for its researchers. — via 1 2
4. Perplexity Computer rolled out Ultrafast versions of GPT-6 Astra and SOL 6.1 for Max subscribers, generating tokens faster to complete long-running Computer tasks sooner; users select Astra or SOL 6.1 in Custom and set speed to Ultrafast, with a warning that it consumes quota faster. Computer also integrated the Polymarket API, letting queries mention or @Polymarket to pull market-implied probabilities, volume, and liquidity data from more than 250,000 active prediction markets for market and real-time trend research. — via 1 2 3

3. Research, Funding, and Industry Signals

1. Anthropic opened applications for two new Fellowships: STEM Fellows in science and engineering, and Conceptual Reasoning Fellows in philosophy, cognitive science, economics, policy, and law. Selected fellows will work at Anthropic for four months starting January 2027, with the goal of helping Claude reason well about the hardest problems in those fields; applications close October 28. — via 1
2. NVIDIA announced new investments in Boltz Bio, Foresight Works, Hiverge AI, Prima Mente, and Riverlane in London, spanning drug discovery to quantum computing. NVIDIA also said the Nordics are building on its technology, with Spotify using it to improve podcast recommendation efficiency, ABB Robotics training and testing robots in simulation, and startups applying open models to Swedish transcription and patent research. — via 1 2
3. Mistral AI is partnering with CERN to explore large-scale AI models in particle physics, aiming to train a single "large physics model" on billions of simulated collision events. — via 1
4. Ethan Mollick argued that the larger the gap between AI capability expansion and actual usage (the "overhang"), the more enterprises pay for locking in "good enough" solutions based on current AI capability ceilings, so they should ensure they can benefit from continued investment in stronger AI by major labs. He also said chain-of-thought — even summarized CoT — has suddenly lost value in the agent era, changing how people collaborate with AI: to guide AI effectively and gain observability, users now need to request intermediate outputs and tests and use AI to investigate discovered problems. — via 1 2
5. Kai-Fu Lee said hiring a PhD with an IQ of 300 and assigning them to customer service wastes talent, and by analogy CEOs must personally drive AI transformation, with the core being to rethink where AI can create the most value rather than just improving workflows. — via 1