NVIDIA open-sources Nemotron 3.5 Lightning as ChatGPT and Gemini both pass 1 billion users

August 11 (PT) was a dense day for model releases and scale milestones. NVIDIA open-sourced Nemotron 3.5 Lightning, a 30B MoE model aimed at always-on agent workloads with high throughput and low cost; SGLang and Cline followed within hours. Meta returned to open weights with Muse Glimmer, a 30B dense model. In the same window, ChatGPT and Gemini both announced more than 1 billion monthly users. OpenAI put out several announcements at once: a Linux desktop client, ad testing in ChatGPT, and the departure of COO Brad Lightcap. On the security side, a research team claimed it could read the encrypted reasoning of frontier models through an API vulnerability, which in turn stoked the debate over text watermarking and EU compliance. The day was dominated by product launches, open-weights releases, and scale competition, with research progress running in parallel.

One: NVIDIA open-sources Nemotron 3.5 Lightning for always-on agent workloads

NVIDIA released Nemotron 3.5 Lightning, a customizable open model built for always-on agents. It is a mixture-of-experts model with 30B total parameters and 3B active parameters, distilled from NVIDIA’s frontier Nemotron 3 Ultra. It supports a context window of up to 1M tokens, and weights are available on Hugging Face in BF16 and NVFP4 precision, able to run locally on RTX PCs, DGX Spark, and Jetson devices.

NVIDIA’s claimed figures: up to 4x higher token-generation throughput and roughly 30% shorter task-completion times than comparable open models. Cline’s blog cites its preliminary results: PinchBench 86.2, SWE-Bench Verified 54.3, and 73 on AA-Omniscience non-hallucination. The model supports three speculative-decoding techniques — MTP, DFlash, and DSpark — and can be reached through an OpenAI-compatible API for agent workflows. NVIDIA’s Bryan Catanzaro noted it shares the 3.0 Nano architecture, adds speculative decoding, and matches the intelligence of 3.0 Super.

Ecosystem adoption was immediate: SGLang announced day-0 support, Cline integrated it for free the same day, and Perplexity listed it on the Agent API at $0.0115 per 1M input tokens and $0.17 per 1M output tokens. Jensen Huang personally promoted it, calling it “Lightning strikes for continuous and long-run agents.” This is NVIDIA’s latest attempt to package open small models with high-throughput inference into agent workflows.

Evidence boundary: the 4x throughput and 30% time-reduction figures are NVIDIA’s own claims; the benchmark scores come secondhand via Cline’s blog and have not been independently replicated yet.

Sources:

Two: Meta returns to open weights with Muse Glimmer

Meta released Muse Glimmer, a 30B multimodal reasoning model with open weights. Sebastian Raschka’s architecture teardown notes it is a dense model (not MoE) with a 131k context window, hybrid attention mixing grouped-query attention (GQA) and sliding-window attention (SWA) at a 3:1 ratio, plus the gated attention that has become common recently. Its standout feature is KV-cache efficiency: about 52 KiB per token in BF16, well below Qwen3.6 27B (~64 KiB) and Gemma 4 31B (~840 KiB), helped by an extreme 32 query heads / 2 KV heads configuration. The model is distilled from Muse Spark, which remains API-only for now.

Meta’s own benchmarks show Glimmer mostly ahead of Qwen3.6, while the independent Artificial Analysis Intelligence Index places it slightly behind — Raschka’s verdict: “a few days of using it will tell where it really ranks.” US Treasury Secretary Bessent publicly welcomed the release as “another win for American innovation,” pulling open weights back into policy discourse. For the community, Meta reopening its weights matters more than the single model — it has been a long time since the Llama days.

Evidence boundary: architecture and KV-efficiency figures come from a single teardown, and independent rankings are still early.

Sources:

Three: Researchers claim they can read “encrypted reasoning”; watermarking and compliance heat up

A research group (kotekjedi_ml / Panfilov team) claims it found a vulnerability in the APIs of OpenAI, Anthropic, Google, and other major providers that lets them read the encrypted thinking of reasoning models. Per The Decoder, scanning roughly 7,000 public conversations surfaced 62 API keys, 33 email addresses, and 33 passwords; via jailbreaking, Anthropic’s Haiku 4.5 could transcribe Opus 4.8’s raw reasoning verbatim; decoding 10,000 reasoning traces cost about $720 in API fees. The claim spread widely, retweeted by Yarin Gal, Stanford NLP, and others — one of the most-shared security stories of the day.

Text watermarking dominated the follow-on discussion. In an updated user note from about ten days earlier, OpenAI stated that OpenAI, Anthropic, Google, Meta, Microsoft, and Mistral have all signed the EU Code of Practice on Transparency in AI-Generated Content, committing to machine-readable marking of AI outputs including text; xAI has not signed. OpenAI said its next step is extending provenance signals to text. In the Chinese community, a widely shared explainer walked through how watermarking works: at generation time a key splits the vocabulary into two groups and quietly biases token choice toward one group; detection re-derives the split and checks whether the green-group share significantly exceeds 50%. Rewriting defeats most schemes — even Google’s SynthID can be bypassed with free rewriting tools — and labs do not expect watermarks to stop determined evaders; the effort is largely compliance-driven.

Evidence boundary: the reasoning-extraction vulnerability is a single team’s claim, not yet independently reproduced; watermark capabilities and bypasses are drawn from researcher and community accounts.

Sources:

Four: The general-agent war begins: Grok Bot launches, ChatGPT/Codex lands on Linux

The day’s most active product line was general-purpose agents. SpaceXAI launched Grok Bot (early beta): each bot gets its own cloud computer, can sign in to Gmail, Salesforce, LinkedIn, Zendesk, and other tools to operate real interfaces, and delivers finished work rather than instructions. It runs asynchronously 24/7, supports multiple parallel bots that share context and hand off tasks, and can record a user-demonstrated workflow as a reusable Routine. Per an earlier The Information report, this is Cursor’s next-generation general agent, codenamed Sand, renamed Grok Bot after SpaceX’s acquisition of Cursor — in effect Cursor’s agent harness expanding from coding to all knowledge work. Grok Bot is available to SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium users, and xAI’s macOS download link points directly to Cursor. Meanwhile, reports said Grok 4.6 would ship this week (delayed yet again).

OpenAI announced the same day that ChatGPT desktop is now on Linux (Preview), bundling ChatGPT, Work, and Codex with continuity across Projects, local dev environments, and browser workflows. Tibo celebrated: “you can cancel that MacBook order.” On the Chinese side, DeepSeek registered a WeChat account for a “DeepSeek Harness” team with beta invites reportedly sent since August 1, widely read as a sign of an official domestic coding agent to rival Claude Code/Codex — still rumor, not confirmed. Separately, Z.ai reset quotas for all Coding plan subscribers to celebrate Zcode passing one million users, turning founder “Tibo” into a community meme.

Evidence boundary: Grok Bot details come from community write-ups and the official site; DeepSeek Harness rests on a WeChat account registration and developer emails, without official confirmation.

Sources:

Five: OpenAI churn and commercialization: Lightcap out, ads in testing, Daybreak on AWS

OpenAI COO Brad Lightcap announced he is leaving to “start something new.” He joined in 2018, became COO in 2022 and ran the OpenAI Startup Fund, and was effectively the commercial No. 2 by March 2025, responsible for daily operations, global expansion, business strategy, and infrastructure. This April he stepped down as COO to run Special Projects reporting directly to Sam Altman; four months later he departed. Community tallies count 16 senior-role changes at OpenAI in the past 12 months, including COO, CEO of AGI Deployment, Chief Futurist, and Head of Sora. The timing — after a $122B raise at an $85.2B valuation, ahead of a possible IPO — has invited considerable speculation.

On products and commercialization, the OpenAI blog posted two items the same day: it is beginning to test ads in ChatGPT to support free access (with clear labeling, answer independence, privacy protections, and user control, per the announcement), and Daybreak cybersecurity models are now available on Amazon Bedrock for enterprise security workflows. Separately, AI Valley reported the next flagship, codenamed Astra (possibly GPT-6), could launch this month at a rumored ~10 trillion parameters; a release was reportedly paused the previous day over security concerns, with Astra designated OpenAI’s first “critical” cybersecurity-capable AI. A larger model codenamed Doug is also in preparation. The Verge reported that the unreleased Astra solved 10 long-open math problems — covering sphere packing, error-correcting codes, and the existence of non-sofic groups — backed by a 250+ page paper and Lean formal proofs.

Evidence boundary: Lightcap’s departure comes from his own statement and community recaps; Astra’s parameter count and Doug are unconfirmed reports; the math results rest on OpenAI’s paper and The Verge’s coverage.

Sources:

Six: ChatGPT and Gemini both pass 1 billion monthly users

OpenAI and Google confirmed on the same day that their chatbots have crossed the 1 billion monthly user mark. OpenAI disclosed in an August 6 blog post that ChatGPT has over 1 billion monthly active users and said weekly users had already reached 1 billion in July; Google CEO Sundar Pichai announced Gemini has 1 billion monthly users — Google’s fastest-growing product ever and its 14th to reach the milestone — up from 750 million in February. Demis Hassabis and the official Gemini account both celebrated publicly. Some in the community argued a meaningful share of Gemini’s growth comes from embedding across Google apps rather than the standalone app’s own pull; fair as far as it goes, but both companies now reporting “1 billion MAU” as a public number is itself a significant marker for the scaled phase of conversational AI.

Sources:

https://www.theverge.com/ai-artificial-intelligence/978113/chatgpt-gemini-1-billion-users

Seven: Video generation, three releases: Seedance 2.5, LTX-2.5, and FLUX 3 Video

Video generation saw the most product churn of the day. Runway shipped Seedance 2.5, supporting 50 unique character references and clips up to 30 seconds synced to music. A Chinese community tester shared three signals for spotting spliced AI long takes (brightness ramps at segment starts, action resets at seams, and breaks in secondary motion), and noted Seedance 2.5 is available unlimited for up to 33 days on Higgsfield.

Lightricks released LTX-2.5, its next-generation video model, summarized by the company as quality, continuity, control, and efficiency. It introduces Diffusion Fidelity Rendering, which allocates compute by scene complexity, and a new Diffusion Video Decoder that improves faces, text, and fast motion. Native Multi-shot generates multiple continuous shots in one pass while preserving character, environment, lighting, and audio consistency across cuts. It pairs a Gemma 4 12B text encoder with an optional Prompt Enhancer and can decide generation length from the described action. Pro features include native 4K HDR, EXR input/output, and RAW workflows. The weights stay open, with a 16GB minimum VRAM for deployment; officially, two GB200s generate a 10-second video in about 6.8 seconds. Black Forest Labs, meanwhile, said FLUX 3 Video ranks No. 2 in the world (16 points off the top) and made it free in its playground until Sunday.

Evidence boundary: LTX-2.5’s 6.8s/10s figure is an official test claim; FLUX 3’s “No. 2” ranking is the vendor’s own statement; the Seedance long-take detection method comes from a single community field test.

Sources:

Eight: Chinese models and founders in motion: Qwen 3.8 near open-source, Ling-3.0-tiny, Lin Junyang’s new lab

Two threads ran on the Chinese side: releases and a founder move. Ant Group’s Bailing open-sourced Ling-3.0-tiny, a native hybrid-reasoning model with 7.9B total and 1.3B active parameters, offered in BF16, FP8, and INT4. Alibaba’s Qwen team released Qwen-MM-Plugins, packaging multimodal abilities — image/video/document reading, audio, and 3D modeling — as Skills plus MCP servers that plug into any agent harness such as Claude Code or Codex; community demos paired it with DeepSeek as “one brain, one pair of hands.” Multiple sources said Qwen3.8 is about a day from open-sourcing, with the dense Qwen3.8-27B due this week (the official word is “this week”).

On people: Lin Junyang, former head of Qwen, officially announced his startup — Pragmatik Labs (p7k) in Shanghai — researching next-generation agents spanning the digital and physical worlds (digital agents plus embodied intelligence). Per The Information, the round is co-led by Gaorong Ventures and Sequoia China with Tencent and the Shanghai Future Industries Fund following, at several hundred million dollars and a post-money valuation around $2 billion. It is the latest in a series of model leaders leaving to build agents.

Evidence boundary: the Qwen3.8 release timing is community-reported (“this week”); Pragmatik’s funding and valuation come from The Information without official company confirmation.

Sources:

High-value briefs

Google AMIE medical video consultations: Google Research and DeepMind’s AMIE system demonstrated real-time clinical video consultation for the first time, built on Gemini and Project Astra; it reads visual and auditory cues, guides a virtual physical exam, and reasons diagnostically in real time. In a randomized study, clinical evaluators rated history-taking and diagnostic accuracy positively, and patient actors preferred the video experience. Research-stage result. https://blog.google/innovation-and-ai/models-and-research/google-research/amie-video-consultations
LMSYS Unified Radix Cache: a single token-keyed radix topology to unify prefix caching across FULL, SWA, and MAMBA components of hybrid models, reusing KV for shared prefixes across attention components — aimed directly at hybrid-architecture inference cost. https://www.lmsys.org/blog/2026-08-11-unified-radix-cache
Metal compatibility layer for macOS VMs: researchers built a process-level compatibility layer for Metal capability queries inside macOS VMs so llama.cpp can pick newer Metal kernels; on M1 Ultra, TinyLlama 1.1B prompt processing sped up 11.08x and token generation 16.36x (near 98% of bare metal), while Gemma 4 12B saw 7.20x and 14.54x. https://github.com/trycua/cua/blob/main/blog/gpu-passthrough-macos-vms.md
BDH-CQ latent-space reasoning: scored 29.5% on ARC-AGI 1 at roughly $0.0007 per task, reasoning recurrently in latent space rather than via chain-of-thought, with Transformer-like scaling verified up to 600B parameters. https://x.com/omarsar0/status/2087191397111656738
TwiL-LM3 (3B): webAI Intelligence Lab open-sourced a 3B formal-reasoning model that beats OpenAI GPT-OSS-120B on 4 of 5 formal-reasoning benchmarks — another data point for small, specialized models. https://x.com/shao__meng/status/2087167721465155962
ARC-AGI progress comparison: a widely shared post contrasts o3 (April 2025, 6.5% on ARC-AGI-2) with GPT-5.6 Sol Max (92.5% after 449 days), while per-task reasoning cost rose only from $0.834 to $1.44 — used as an argument that model capability is still advancing fast. https://x.com/MaxForAI/status/2087124713130557757
Stanford’s 29-page agent memory guide: argues the real test of agent memory is using it to drive the next action, not recalling what was seen; models that nearly max out LoCoMo fall apart when memory must drive a decision. https://x.com/Mnilax/status/2087257363594051818
Cloudflare H1 DDoS report: a 519% surge in hyper-volumetric DDoS attacks detected in the first half of 2026, with the report tying growth to geopolitical conflict. https://x.com/Cloudflare/status/2087163380666601528
Local deployment payback math: OpenCode figures show an average Go user spent about $1.14/day on DeepSeek V4 Flash over the past week; a dual-DGX setup costs about $10K, implying ~24 years to break even at the same usage, or 2.4 years at 10x usage. HN consensus: local setups mostly buy privacy, offline access, and fixed models; pure cost savings fit only a few high-utilization scenarios.
Benchmark vs. real-codebase gap: Qodo’s AI Code Review Academy cites a 2025 study where the same model scored 84-89% on an isolated benchmark but 25-34% inside a real codebase; it recommends testing review tools on 10-20 of your own PRs. https://x.com/omarsar0/status/2087183764057460972
AI companies bulk-buying books (rumor): a community post claims AI companies are purchasing, scanning, and destroying physical books by the millions (pre-2022 print editions) to manage copyright exposure — single-source and unverified. https://x.com/KanikaBK/status/2087187528160022720
WeChat Moments “AI help write” in gray release: screenshots show WeChat’s Moments experimenting with an “AI help write” feature that drafts post copy from an uploaded photo — a multimodal entry point reaching a mass-market product, worth watching.

🕐 Selected hourly signals

PT time	Signal	Why it matters
03:00	Grok Bot launch write-ups flooded feeds, with parallel teardowns of its cloud computer, multi-bot collaboration, and Routines	Core event reshaping the general-agent landscape
02:00	Brad Lightcap’s departure spread through the Chinese community, framed as pre-IPO leadership churn	High-signal personnel move coinciding with OpenAI’s IPO timeline
22:00	A detailed explainer of text watermarking (green/red group hash mechanism and rewrite bypasses)	Moved the watermark debate from emotion to technical detail
21:00	Side-by-side ARC-AGI progress: o3 vs GPT-5.6 Sol Max	Data for the “is AI stagnating” argument
21:00	NVIDIA official accounts amplified Nemotron 3.5 Lightning and local-deployment content	Official framing cross-checked against third-party takes
20:00	Z.ai resets quotas for all Coding plan subscribers as Zcode passes 1M users	Escalating subscription competition in coding agents
17:00	Community H3 video-workflow tests: 8V8A output settings, CK Attention ~12.1% faster, low-VRAM options	Translates video capability claims into concrete engineering parameters
08:00	Local-deployment payback math ($1.14/day vs. $10K dual DGX) circulated on HN and Chinese feeds	A reality check on the “everyone should own compute” narrative
08:00	OpenAI’s watermark note updated; six labs signed the EU code, xAI did not	Direct evidence of compliance boundaries and vendor positions

Editorial conclusion

The day’s through-line is clear: frontier models continue on the twin tracks of open weights plus ecosystem adoption (Nemotron, Muse Glimmer), chat products enter a 10-billion-user scale race, and general-agent platforms (Grok Bot, ChatGPT on Linux, the DeepSeek Harness rumor) all moved within the same week. On security, the claimed ability to read reasoning traces and the text-watermarking compliance push put “can we trust model outputs” back on the agenda. For readers, the things worth tracking are not the individual launches but three questions: whether open-model inference costs really fall as NVIDIA claims, whether general-agent platforms can move beyond coding, and what form watermarking and reasoning protection take under regulatory pressure.

Sources and method

This daily is based on 30 raw files in the 2026-08-11 PT archive (20 hourly captures, 6 substantive named sources, and a status file); the signal pool is rated rich. Claude, Google Research, and Chrome official sources had no new posts that day, and XiaoHu.AI failed to capture; remaining sources cover model releases, product updates, security research, open-source projects, and community field tests. Vendor self-reported data and single-post views are flagged as such in the body.

One: NVIDIA open-sources Nemotron 3.5 Lightning for always-on agent workloads

Two: Meta returns to open weights with Muse Glimmer

Three: Researchers claim they can read “encrypted reasoning”; watermarking and compliance heat up

Four: The general-agent war begins: Grok Bot launches, ChatGPT/Codex lands on Linux

Five: OpenAI churn and commercialization: Lightcap out, ads in testing, Daybreak on AWS

Six: ChatGPT and Gemini both pass 1 billion monthly users

Seven: Video generation, three releases: Seedance 2.5, LTX-2.5, and FLUX 3 Video

Eight: Chinese models and founders in motion: Qwen 3.8 near open-source, Ling-3.0-tiny, Lin Junyang’s new lab

High-value briefs

🕐 Selected hourly signals

Editorial conclusion

Sources and method

智简 Smart&Concise