By the Numbers
Statistically Speaking
54
on the Artificial Analysis Intelligence Index (a
Google put its new Gemini 4 Hydrogen model into Search for e
24%
less of a subscriber's allowance on Google's
Google put its new Gemini 4 Hydrogen model into Search for e
One Thing to Tell Your Friends
One Thing to Tell Your Friends
Jensen Huang, the only boss Nvidia has had in its 33 years, resigned after a scandal involving a woman US officials suspect of spying for China, and Nvidia's stock briefly fell about 14%.
Summary
TL;DR
Top Stories
Trends
AI's suppliers are cashing in while builders hit snags, Open, and AI agents are being sent out to run errands.
Creative AI
Dev Tools
Research
Business
Education
Surprising
Worth Watching
GitHub
HuggingFace
Leading models: ricursiveAI/Ripple-65B (7,667), thinkingmachines/Inkling-2 (12), and meituan-longcat/LongCat (710).
Product Hunt
API Pricing
One new model, no price changes: Google made Gemini 4 Hydrogen generally available on November 12 at $1.50 per million input tokens and $7.50 per million output tokens, with a free tier as well; Google's pricing page does not list its context size, though its announcement says 1M tokens.
arXiv
AI Isn't Getting Dumber, We're Getting Pickier — For the very same unchanged models, the share of votes saying "both answers are bad" rose by about 1.7 percentage points a month, and in 14 frequently compared pairs it climbed in every one, from 15.5% to 23.3%, while which model won stayed steady.
FYI
Hot off the Presses
01
Nvidia CEO Jensen Huang resigned, and finance chief Colette Kress is taking over for now
What this means for you: Nvidia makes the chips behind almost every major AI service and sits in many retirement funds, so its leadership vacuum could ripple into AI prices and your investments.
Previously: November 11 - Nvidia's board was reportedly preparing to remove Huang.
Today: Nvidia announced that Huang resigned as CEO and board member, effective immediately. Its statement gives no reason. News reports tie his exit to admitted affairs, including one with an executive at chip buyer Megaspeed whom US officials suspect of spying for China.
“Nvidia shares fell as much as 14% before closing down about 4%”
- Shares fell as much as about 14% in early trading - then recovered to close down about 4.3%, per CNBC
- Market value dipped below $4 trillion during trading - per Bloomberg, before closing just above it per CNBC, with AI cloud renter CoreWeave down about 12% (Nebius fell 17%, though it also cut its own forecast that day)
- Nvidia now has its first-ever board chair - Harvey C. Jones, a former Synopsys CEO, a split observers read as the board taking firmer control
- Huang agreed to testify under oath before Thanksgiving - in a closed-door session, said House Foreign Affairs chair Brian Mast
- More legal pressure is building - Semafor reports Justice Department investigations could sweep up Nvidia, though it saw no sign of criminal charges
02
Google put its new Gemini 4 Hydrogen model into Search for everyone
What this means for you: Google Search answers should handle long, specific and typo-filled questions better starting today, at no cost to you.
Hydrogen is Google's new fast, everyday model, positioned below Google's frontier Gemini 4 Carbon and Argon models (covered November 10). Google calls it "the biggest model upgrade to Search yet." It is also rolling out in the Gemini app, in Google Sheets and to developers.
- Nearly ties the best mid-priced rivals - it scores 54 on the Artificial Analysis Intelligence Index (a combined test of reasoning, coding and knowledge), versus 55 for Claude Sonnet 5.5 and GPT-6.1 Sol
- Costs about a third as much as Sonnet to run that test - per Artificial Analysis
- $1.50 per million input tokens - and $7.50 per million output tokens for developers (a token is roughly three-quarters of a word)
- Google says it makes up facts over 5 times less often - compared with its older Gemini 3.8 Flash
- A new "Orchestrator" teams it with Carbon for coding - using about 24% less of a subscriber's allowance on Google's tests
03
Washington is moving to close the AI chip "rental" loophole
What this means for you: The fight over who gets America's best AI chips is heating up fast, and its outcome will shape how powerful foreign AI rivals become.
US rules bar sales of Nvidia's most advanced chips to China, but Chinese firms can legally rent them in data centers abroad. Huang was the loudest industry voice against tighter rules. Since he left, lawmakers and AI companies have moved quickly.
- OpenAI now backs three chip-control bills - the Chip Security Act, the MATCH Act and the AI OVERWATCH Act, per Bloomberg and Punchbowl News, reversing its stance from about three weeks ago
- The Commerce Department is finishing a rental rule - it would require licenses for cloud providers in seven countries, including Singapore and Malaysia, per The Information
- Chinese AI firms are rushing to lock in overseas capacity - before the rule lands, Reuters reports, citing three people
- Senate Majority Leader John Thune - says there is a "high probability" Congress adds chip restrictions; the year-end defense bill is the likely vehicle, per the FT
04
Google and Samsung's AI glasses are up for preorder
What this means for you: Gemini-powered glasses that look like regular frames go on sale December 10, so expect to see them as a holiday gift option.
Previously: November 1 - Samsung's first AI glasses were expected in mid-December.
Today: Samsung builds the hardware, Google supplies the software, and eyewear brands Warby Parker and Gentle Monster design and sell the frames. There is no screen in the lenses. You talk to Gemini, and it can see through the camera to answer questions about what you are looking at.
- From $395 at Warby Parker and $449 at Gentle Monster - with prescription lenses available
- About 41 grams with a nine-hour battery - roughly 10 grams lighter than Meta's latest glasses
- Privacy guardrails - recording stops if the camera light is covered, and bystanders' faces are blurred before Gemini sees an image
- Narrow phone support for now - Pixel 7 or newer, recent Galaxy phones, or iPhone 13 Pro or newer
05
AI chatbots are sending shoppers to risky websites
What this means for you: Before you buy from a store a chatbot recommends this holiday season, check the web address yourself.
Brand-protection firm Podqi asked ChatGPT, Gemini and Perplexity where to buy products from 151 brands, including Lululemon, Birkenstock and Le Creuset. Across 6,795 answers, the bots surfaced knockoff stores, old brand web addresses now owned by strangers, and dead addresses anyone could buy.
“Perplexity pointed shoppers to risky websites in 15% of answers”
- Perplexity surfaced risky sites in 15% of answers - ChatGPT in 8% and Gemini in 4%
- Smaller brands were hit about twice as often - as large companies
- Questions about country-specific stores were the worst - the bots sometimes invented regional sites that do not exist
- Warnings were rare - ChatGPT flagged the risky site only 24% of the time it showed one
Trends & Themes
Trends & Themes
AI's suppliers are cashing in while builders hit snags
Why this matters to you: The companies selling parts for AI are thriving, but delays at the data centers themselves show the boom is not frictionless.
The money is real, but turning it into working data centers depends on power, permits and partners. Watch for more forecast cuts tied to delayed sites.
- Networking giant Cisco expects about $9 billion - in AI sales to big cloud companies this fiscal year, up from its earlier $7.5 billion forecast
- AI chipmaker Cerebras grew revenue 70% - and forecast well above expectations, sending its stock up 17% after hours; it also has a new Microsoft deal signed in October
- AI cloud company Nebius cut its 2026 forecast by $200 million - blaming delays at a partner's New Jersey site, and its shares fell 17%
- SpaceX plans more than $74 billion of AI campuses - in Texas and Arizona, but the Arizona site still needs local approval
Open-source AI is looking for a new champion
Why this matters to you: Free AI models you can download keep prices down and keep AI out of a few companies' hands, and their biggest corporate backer just lost its leader.
An Nvidia insider told TechCrunch work on its free models will continue. Free models help Nvidia sell chips, so the business case stays even without Huang.
- Huang personally drove Nvidia's free Nemotron models - and organized a letter from 25 companies urging the US not to over-restrict open models, per TechCrunch
- Nvidia's $12.9 billion purchase of Hugging Face is not finished - Hugging Face (the main website for sharing free AI models) says it will fight for open source "harder than ever"
- Researcher Nathan Lambert called Huang's exit - "very very bad for the open model cause in the US"
- LibreAGI, an anonymously run members-only file-sharing tracker - keeps backup copies of free AI models in case hosting sites remove them
AI agents are being sent out to run errands
Why this matters to you: AI helpers are starting to call, email and message real businesses and people for you.
Businesses on the receiving end are not all welcoming. Many receptionists hang up once they realize an AI is calling.
- Meta's Muse assistant can now phone US businesses - for paying users, from $20 a month, to deal with cable companies or restaurants
- Nectar's new agent finds social media creators and messages them - drafting outreach in a brand's voice across a pool of over 300 million profiles
- Note-taking app Granola now writes a nightly "Agent Brief" - about you, which any of your AI agents can read, per Ben's Bites
- Amazon upgraded AgentCore - the service companies use to run agents, with sessions starting in under 2 seconds
Smaller, specialized AI is moving inside everyday software
Why this matters to you: Many AI features you use will soon run on cheap, fast models, or on your own device, instead of a giant chatbot.
The pattern is a division of labor. Small, fast models handle routine choices, and expensive models are saved for hard thinking.
- Companies like Glean, UiPath and Workday - showed apps using fast "decision models" that pick options instead of writing text, at a developer conference covered by Latent Space
- A hobby project runs a 16-billion-part AI model on a TV box - with about 1 gigabyte of usable memory
- IBM released small 3B and 8B models - with a new labeling format that helps them ignore hidden instructions in documents
- Apple's Foundation Models framework now accepts any AI model - including Claude and Gemini, through the same interface as Apple's on-device model
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Business & Industry
Business & Industry
Education
GenAI in Education
Surprising
Surprising & Under-the-Radar
Worth Watching
Signals to Track
01
Tencent may move half its staff to AI-first products
One of China's biggest tech companies is willing to redeploy roughly 50,000 people.
Founder Pony Ma said in his annual speech he is willing to deploy half of Tencent's roughly 100,000 workers to AI-native businesses, Bloomberg reports, citing sources. Earlier this year he admitted Tencent had moved too slowly on AI. If it happens, WeChat's more than a billion users could see AI woven through apps they use daily.
02
China's biggest shopping day shows what AI really sells
AI shopping helpers boomed during Singles Day, but overall spending barely moved.
JD.com's Dongdong assistant drew nearly 80 million users seeking advice, more than triple last year, and reports say AI agents were best at upselling services like house cleaning and travel. Yet early data showed festival sales up just 2.4%. If that holds, AI shopping helpers may shift where money goes more than how much people spend.
03
A laptop-sized check for engineered germs
A startup wants to screen DNA for dangerous pathogens in the field, with no internet connection.
Radical Numerics' Omnii Defense pairs a laptop with a USB gene sequencer and scores a sample from 0 to 1 for threat, including whether it looks AI-designed. The company reports 95.8% recall on known pathogens, but independent testing is still pending. If it works, health officials could check outbreaks on site instead of waiting for a lab.
04
Meta handed its AI infrastructure to an Amazon cloud veteran
Meta's longtime infrastructure chief is stepping down mid-buildout.
Santosh Janardhan will leave in early 2027, and Dave Brown, who spent about 19 years at Amazon running its core cloud computers, takes over. Observers read it as Meta's bottleneck shifting from designing hardware to buying and running capacity. That could decide how fast Meta's AI features reach Facebook, Instagram and WhatsApp.
GitHub Trending
Top Repos Today
A self-hosted workspace where AI "coworkers" each get their own computer led the new AI projects on GitHub today, adding nearly 300 stars. GitHub's main trending page listed only 13 projects, so we also drew on its Python, TypeScript, Go and Rust trending lists. We skipped projects featured in the past week that are still trending, including Google Chrome's Modern Web Guidance (covered November 10, up 739 stars today), OpenAI's Codex Security tool, Meta's Astryx, Addy Osmani's agent skills, free-claude-code (yesterday's #1), Vercel's agent-browser, Multica, the Home Assistant connector and the stock-analysis bot. We also left out a "jailbreak" chat app, an "uncensored" AI studio and a research skill pitched for digging into individual people. Today's leader is CopilotKit's OpenBot, which added 298 stars.
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +298 · 📦 Total: 11,589
📜 License: MIT · 👤 By: CopilotKit (company)
🎯 Time to value: 20 minutes
📜 License: MIT · 👤 By: CopilotKit (company)
🎯 Time to value: 20 minutes
What it is: A self-hostable workspace for "AI coworkers," where each agent gets its own computer with a browser, files and tools. Every action is checked against your rules before it happens and logged afterward, and it works with OpenAI, Claude, Gemini or open models. Why you'd want it: Give a team AI helpers that can do real computer work while keeping sign-in, permissions and an audit trail under your control.
| ✓ Pros | ✗ Cons |
|---|---|
| Free desktop app for Mac and Windows | Business features such as company sign-in sit in an early-access tier |
| Bring your own models and existing agents | Running agents with their own computers takes real resources |
| Permissive license allows rebranding | Another platform for a team to learn and maintain |
GitHub - CopilotKit/OpenBot: Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent.
Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent. - CopilotKit/OpenBot
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +297 · 📦 Total: 25,698
📜 License: Apache-2.0 · 👤 By: NVIDIA (organization)
🎯 Time to value: 10 minutes
📜 License: Apache-2.0 · 👤 By: NVIDIA (organization)
🎯 Time to value: 10 minutes
What it is: A safety checker that scans add-on skills for Claude Code, Codex and similar agents before you install them. It looks for 71 warning patterns, such as hidden instructions (prompt injection) or code that sends your data elsewhere, and gives each skill a 0 to 100 risk score. NVIDIA says 26.1% of the 31,132 skills in its research set had at least one weakness. Why you'd want it: A quick "is this safe?" check before letting a stranger's add-on run with your agent's access.
| ✓ Pros | ✗ Cons |
|---|---|
| Scans Git repos, links, zip files or single files | The optional deeper AI check costs model time |
| Reports work in common security formats | A clean score is not a guarantee of safety |
| Free and from a major company | Needs a Python setup on the command line |
GitHub - NVIDIA/SkillSpector: Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, security risks, prompt injection, data exfiltration, and supply-chain risks in Claude Code, Codex, and MCP skills before you install them.
Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, security risks, prompt injection, data exfiltration, and supply-chain risks in Claude Code, Codex, and MCP skills b…
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +184 · 📦 Total: 539
📜 License: MIT (per the project's README; GitHub does not detect it) · 👤 By: Tencent (organization)
🎯 Time to value: 30 minutes
📜 License: MIT (per the project's README; GitHub does not detect it) · 👤 By: Tencent (organization)
🎯 Time to value: 30 minutes
What it is: A self-hosted workspace where coding agents such as Claude Code, Codex and Tencent's CodeBuddy follow a fixed procedure for every job: plan, build, review and test, each handled by a separate role. Each run happens in its own sealed-off container, pauses at human approval steps and keeps a full log. Why you'd want it: Get the same careful process from AI coding agents every time, instead of results that depend on the luck of a single run.
| ✓ Pros | ✗ Cons |
|---|---|
| Ready-made procedures for common jobs such as bug fixes | Needs Docker and a database server to run |
| Human sign-off steps where they matter | Very new, with a small user base so far |
| Demo project runs without any AI key | License file is not recognized by GitHub, so check before company use |
GitHub - Tencent/LoopForge: Self-hosted workflows for coding-agent delivery. Orchestrate CodeBuddy, Codex and Claude Code with scripts and human approval gates. Includes Local workflows for coding assistants.
Self-hosted workflows for coding-agent delivery. Orchestrate CodeBuddy, Codex and Claude Code with scripts and human approval gates. Includes Local workflows for coding assistants. - Tencent/LoopForge
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +176 · 📦 Total: 134,546
📜 License: MIT · 👤 By: ggml-org (organization)
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: ggml-org (organization)
🎯 Time to value: 15 minutes
What it is: The widely used free engine for running AI language models on your own computer, from laptops to graphics-card servers. Many popular desktop AI apps run on it behind the scenes, and it supports most open models soon after they come out. Why you'd want it: Run open AI models privately and offline, with no per-use fees.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on Mac, Windows, Linux and many kinds of chips | Command-line tool, less friendly than desktop apps built on it |
| Huge community and fast support for new models | Big models still need a lot of memory |
| Permissive license | Frequent updates can change settings |
GitHub - ggml-org/llama.cpp: LLM inference in C/C++
LLM inference in C/C++. Contribute to ggml-org/llama.cpp development by creating an account on GitHub.
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +151 · 📦 Total: 29,797
📜 License: Apache-2.0 · 👤 By: Qwen team at Alibaba (organization)
🎯 Time to value: 10 minutes
📜 License: Apache-2.0 · 👤 By: Qwen team at Alibaba (organization)
🎯 Time to value: 10 minutes
What it is: An open-source AI coding agent that works in your terminal, code editor, desktop app, browser and even chat apps such as Telegram. It is built around Alibaba's Qwen models but also works with OpenAI, Anthropic, Gemini or local models, and includes helpers like memory, sub-agents and add-on skills. Why you'd want it: A free, open alternative to paid coding agents that lets you pick, or switch, the model underneath.
| ✓ Pros | ✗ Cons |
|---|---|
| Open source, including the Qwen models it is tuned for | Best experience assumes a Qwen account or subscription |
| Many install options, including Homebrew and npm | Fast-moving project, with a separate 1.0 preview version |
| Works with local models for privacy | Quick-install script downloads from Alibaba Cloud servers |
GitHub - QwenLM/qwen-code: An open-source AI coding agent that lives in your terminal.
An open-source AI coding agent that lives in your terminal. - QwenLM/qwen-code
GitHub
HuggingFace Trending
Top Models Today
Most of Hugging Face's top 20 trending models were featured here in the past week, including MiniMax's H3-Image, Nous Research's Arete-27B, Sperid Labs' Idol, Lightricks' LTX-2.5, Qwen3.8-27B, Kimi K3.1 Pro and yesterday's ZONOS3-Mini and Mirage. We skipped those, several "uncensored" copies of other models, a 0.6-billion-parameter retrained voice model and repackaged copies, then looked further down Hugging Face's trending ranking. Today's picks are big and new: a 65-billion-parameter coding model from a five-person team, new giant models from Thinking Machines and Meituan, and a voice model that can hear sarcasm.
A model trained from scratch by a team of five that its makers say beats far larger open models at coding tasks.
📥 Downloads (30d): 7,667 · 📜 License: Apache-2.0
👤 By: Ricursive Intelligence · 🎯 Task: Text generation (coding and agents)
📐 Size: 65B
👤 By: Ricursive Intelligence · 🎯 Task: Text generation (coding and agents)
📐 Size: 65B
What it is: A 65-billion-parameter model trained on 12 trillion tokens of mostly English text and code, with room for about 1 million tokens of input. Ricursive reports a score of 80.52 on TerminalBench 2.1 (a test of agents doing real tasks in a computer terminal), ahead of several open models more than 10 times its size. It uses a plain "dense" design instead of the "mixture of experts" approach most labs use. Why you'd want it: A strong coding agent model small enough to run on a single high-end desktop AI computer such as NVIDIA's DGX Spark, using its compressed version.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license for commercial use | Preview release with light finishing training |
| Compressed version provided for local use | Benchmark scores are the makers' own |
| Long 1M-token input window | Mostly English and code |
ricursiveAI/Ripple-65B-Preview · Hugging Face
Machine learning model with 65B parameters. Download, explore files and run it on Hugging Face.
Thinking Machines' new open model reads text and images and handles very long inputs.
📥 Downloads (30d): 12 · 📜 License: Apache-2.0
👤 By: Thinking Machines · 🎯 Task: Image-text-to-text
📐 Size: 594B total (17B active)
👤 By: Thinking Machines · 🎯 Task: Image-text-to-text
📐 Size: 594B total (17B active)
What it is: A general-purpose model that takes text and pictures and writes text, with an input window of up to 1 million tokens. Despite "Small" in the name, it has 594 billion parameters, but only 17 billion work on each word thanks to a "mixture of experts" design (only part of the model runs at a time). It is also offered through hosts such as Together, Fireworks and Cloudflare. Why you'd want it: A free-to-use frontier-scale model you can customize, or simply call through a cloud provider.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Needs at least 500 GB of graphics-card memory to self-host |
| Available from many cloud providers | Brand new, so download counts are tiny so far |
| Handles images and very long documents | Text-only output |
thinkingmachines/Inkling-2-Small · Hugging Face
Image-Text-to-Text model with 383B parameters. Download, explore files and run it on Hugging Face.
Meituan's giant model built for long, multi-step agent work that mixes coding and seeing.
📥 Downloads (30d): 710 · 📜 License: MIT
👤 By: Meituan LongCat · 🎯 Task: Image-text-to-text
📐 Size: 1.6T total (about 48B active, per Meituan)
👤 By: Meituan LongCat · 🎯 Task: Image-text-to-text
📐 Size: 1.6T total (about 48B active, per Meituan)
What it is: A very large model from Chinese delivery company Meituan that understands images natively and is trained for long agent tasks such as coding, browsing and spreadsheets. It can "think" step by step or answer quickly, and Meituan trained it with safeguards against an AI gaming its own reward scores. Why you'd want it: One of the largest openly licensed models aimed at agents that both read screens and write code.
| ✓ Pros | ✗ Cons |
|---|---|
| Very permissive MIT license | Far too big for anything but data-center hardware |
| Built-in compression for faster serving | Main setup guide targets Chinese AI chips |
| Combines vision and coding in one model | Many scores are Meituan's own |
meituan-longcat/LongCat-2.5 · Hugging Face
Image-Text-to-Text model with 1.8T parameters. Download, explore files and run it on Hugging Face.
An open voice model that talks like a phone call, catching your tone and replying with its own.
📥 Downloads (30d): 0 on this family page (577 for the recommended "Tag" version) · 📜 License: Apache-2.0
👤 By: Adventists AI · 🎯 Task: Speech-to-speech
📐 Size: 32B
👤 By: Adventists AI · 🎯 Task: Speech-to-speech
📐 Size: 32B
What it is: A "full-duplex" voice model, meaning it listens while you talk, replies as soon as you stop and stops if you interrupt. It picks up hesitation, sarcasm and laughter, and can laugh or sigh in reply. The makers report first audio in about 0.25 to 0.6 seconds and up to 40 users at once on one graphics card. Why you'd want it: Build natural-sounding voice assistants or customer-service lines without paying per minute to a cloud voice service.
| ✓ Pros | ✗ Cons |
|---|---|
| Free for commercial use | Needs a large 80 GB graphics card and Linux |
| Can stand in for OpenAI's live-voice connection | Main languages are Chinese and English |
| Three versions trade expressiveness for speed | Benchmark claims are the makers' own |
adventists-ai/Sophon-Realtime-v1 · Hugging Face
Any-to-Any model. Download, explore files and run it on Hugging Face.
A slimmed-down version of MiniMax's video model that makes clips with sound on a gaming graphics card.
📥 Downloads (30d): 326,487 · 📜 License: MiniMax H3 Community License
👤 By: Video Rebirth · 🎯 Task: Image-text-to-video
📐 Size: 13.7B
👤 By: Video Rebirth · 🎯 Task: Image-text-to-video
📐 Size: 13.7B
What it is: A compressed, faster version of MiniMax-H3 that makes 5 to 15 second videos together with a matching stereo soundtrack, from text, a starting picture or up to 12 reference files. A 5-second clip takes about 92 seconds on an RTX 5090 graphics card, and the smallest version is about 9 GB. Why you'd want it: Make short video clips with sound on your own PC instead of paying per clip.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on 24 GB consumer graphics cards | License does not cover the EU, UK, South Korea or the US without MiniMax's approval |
| Ready-made workflows for the ComfyUI app | Still needs large companion files from MiniMax |
| Generates audio and video together | 16 GB graphics cards not yet tested |
VideoRebirthAI/hyperflow-h3-flash · Hugging Face
Image-Text-to-Video model with 14B parameters. Download, explore files and run it on Hugging Face.
#6
IFM/PAN2
A "world model" that predicts, as video, what happens when you take an action.
📥 Downloads (30d): 10 · 📜 License: Apache-2.0
👤 By: Institute of Foundation Models, MBZUAI · 🎯 Task: Image-text-to-video
📐 Size: 29.2B
👤 By: Institute of Foundation Models, MBZUAI · 🎯 Task: Image-text-to-video
📐 Size: 29.2B
What it is: A model that takes a scene (an optional picture plus a description) and a described action, then simulates the outcome as a video of up to 20 seconds. It pairs a language model that reasons about what should happen with a video model that draws it, for uses such as training robots or planning in simulation. Why you'd want it: Researchers and robotics teams can test "what if" scenarios in simulation before trying them in the real world.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Research tool, not a consumer video maker |
| Supported in the popular Diffusers and vLLM tools | Needs data-center graphics cards |
| Videos up to 20 seconds | Very new, with almost no downloads yet |
IFM/PAN2 · Hugging Face
Image-Text-to-Video model with 29B parameters. Download, explore files and run it on Hugging Face.
Product Hunt
AI Launches Today
This week's late AI launches on Product Hunt were about handing whole jobs to agents, from booking sales calls to closing the books. We skipped launches covered in the past week that are still on the leaderboard, including Nod by n8n, CALL-E, Skillbox, Seldon, AppuBot and Almanzor. Upvote counts are Product Hunt's scores when we checked, so final counts may be higher. Alsona, which turns signs that a company is ready to buy into booked sales calls, led with 500 upvotes.
AI agents that turn buying signals into booked sales calls
🔥 Upvotes: 500 · 👤 By: Alsona (launched by co-founder and CEO Jaclyn VC)
💰 Pricing: Free options; flat monthly plan, 30% off for 3 months at launch · 🏷 Category: AI sales tools
💰 Pricing: Free options; flat monthly plan, 30% off for 3 months at launch · 🏷 Category: AI sales tools
Alsona watches more than 32 "buying signals," such as hiring, new funding, technology changes and reviews, and scores companies against the kind of customer you want. Its AI agents then write and send personalized LinkedIn messages and emails to the companies showing interest. The goal is to book a call while the buyer is actively shopping. Verdict: Useful for small sales teams, but automated outreach can feel spammy, so keep a human checking messages.
View on Product Hunt →
AI ad fraud detection that protects your marketing budget
🔥 Upvotes: 303 · 👤 By: ScroogeFrog team (Deepak, Kostia Novofastovskyi, Maria Mashkina, Eugene Chernyak)
💰 Pricing: Free options and paid plans; 75% off the first year at launch · 🏷 Category: Advertising tools
💰 Pricing: Free options and paid plans; 75% off the first year at launch · 🏷 Category: Advertising tools
ScroogeFrog checks your website traffic and Google Ads campaigns for fake clicks, bots and wasted spending, then explains in plain terms what to fix. Agencies and bigger advertisers can get custom fraud-spotting models trained on their own traffic. Verdict: A practical money-saver for anyone running paid ads; check its findings against your own data before cutting campaigns.
View on Product Hunt →
Automate complex accounting work without a single prompt
🔥 Upvotes: 246 · 👤 By: Stefan Kalb (CEO), Jared Kofron (CTO) and team
💰 Pricing: Free plan with no card required; 50% off the Solo plan at launch · 🏷 Category: Accounting software
💰 Pricing: Free plan with no card required; 50% off the Solo plan at launch · 🏷 Category: Accounting software
Instead of typing instructions, you record yourself doing an accounting task on screen while explaining it out loud, including the exceptions nobody wrote down. Latch turns that into a repeatable "sidekick" that does the task the same way each time inside your existing tools. When something doesn't match what it learned, it stops and asks, and every step is logged. Verdict: A smart, show-don't-tell approach for finance teams with repetitive month-end work.
View on Product Hunt →
Read the whole repository with the agent that built it
🔥 Upvotes: 103 · 👤 By: Sarath SP
💰 Pricing: Free and open source (MIT license) · 🏷 Category: Code review tools
💰 Pricing: Free and open source (MIT license) · 🏷 Category: Code review tools
Opens a whole codebase as one big review, so you can comment on any line, file or folder and ask the AI agent why it was written that way. It also reviews your own changes and other people's pull requests (proposed code changes) and posts reviews to GitHub. It runs on your own computer with no account or tracking, and works as a Claude Code add-on. Verdict: A good fit for anyone who has to understand code an AI agent wrote; free and private makes it easy to try.
View on Product Hunt →
Listen to any PDF in a natural voice, then ask Quillo Ai
🔥 Upvotes: 77 · 👤 By: Agâh Emre Sözen
💰 Pricing: Free core app; new accounts get 10 minutes of cloud voice and 10 pages of AI questions, paid prices not listed · 🏷 Category: Productivity
💰 Pricing: Free core app; new accounts get 10 minutes of cloud voice and 10 pages of AI questions, paid prices not listed · 🏷 Category: Productivity
Reads PDFs, Word files, photos and web pages aloud in a natural voice, highlighting each sentence as it goes. You can ask its AI questions about the document, and it includes free PDF tools such as merge, compress and sign on Android, iPhone and the web. Verdict: Handy for commuters and students who would rather listen than read; the free AI allowance is small.
View on Product Hunt →
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Fable 5.2 | $8.00 | $40.00 | 1M tokens |
| Anthropic | Claude Fable 5.1 | $10.00 | $50.00 | 1M tokens |
| Anthropic | Claude Opus 5.5 | $4.00 | $20.00 | 1M tokens |
| Anthropic | Claude Sonnet 5.5 | $2.00 | $10.00 | 1M tokens |
| Anthropic | Claude Haiku 5.5 | $0.10 (up to 100K-token prompts; $0.50 above) | $0.50 ($2.50 above 100K) | 1M tokens |
| OpenAI | GPT-6 Astra | $10.00 (up to 272K-token prompts; $20.00 above) | $50.00 ($75.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Sol | $2.00 (up to 272K; $4.00 above) | $10.00 ($15.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Terra | $1.00 (up to 272K; $2.00 above) | $5.00 ($7.50 above 272K) | 1.05M tokens |
| OpenAI | GPT-6 Luna | $0.10 (up to 272K; $0.20 above) | $0.50 ($0.75 above 272K) | 1.05M tokens |
| Gemini 4 Carbon | $2.00 to Nov 30; $4.00 from Dec 1 | $10.00 to Nov 30; $20.00 from Dec 1 | 1M tokens | |
| Gemini 4 Argon | $2.00 to Nov 30; $4.00 from Dec 1 | $10.00 to Nov 30; $20.00 from Dec 1 | 1M tokens | |
| Gemini 4 Hydrogen (new) | $1.50 | $7.50 | 1M tokens (per Google's announcement) | |
| Gemini 3.8 Flash | $0.75 (to Dec 31; $1.50 from Jan 1) | $3.75 (to Dec 31; $7.50 from Jan 1) | not listed | |
| Groq | GPT-OSS 120B | $0.15 | $0.60 | 131K tokens |
| Groq | Qwen3.8-27B (preview) | $0.80 | $4.00 | 131K tokens |
| Mistral (open weights) | Mistral Large 4 | $0.70 | $2.10 | 1M tokens |
Here is what the major AI companies charge through their Application Programming Interface (API), the paid connection developers use to plug AI models into their own apps. Prices are in US dollars per 1 million tokens (a token is roughly three-quarters of a word).
One new model, no price changes: Google made Gemini 4 Hydrogen generally available on November 12 at $1.50 per million input tokens and $7.50 per million output tokens, with a free tier as well; Google's pricing page does not list its context size, though its announcement says 1M tokens. Every other standard price from Anthropic, OpenAI, Google and Groq matches yesterday's table, and Google's Carbon and Argon prices still double on December 1. Mistral's Large 4 price, OpenAI's context sizes and Google's 1M-token context figures carry over from earlier checks.
What this means: Gemini 4 Hydrogen undercuts Google's own Carbon and Argon models now, and will cost well under half their price once their December increase lands, so teams on Google should test whether it is good enough for their everyday work before December 1.
One new model, no price changes: Google made Gemini 4 Hydrogen generally available on November 12 at $1.50 per million input tokens and $7.50 per million output tokens, with a free tier as well; Google's pricing page does not list its context size, though its announcement says 1M tokens. Every other standard price from Anthropic, OpenAI, Google and Groq matches yesterday's table, and Google's Carbon and Argon prices still double on December 1. Mistral's Large 4 price, OpenAI's context sizes and Google's 1M-token context figures carry over from earlier checks.
What this means: Gemini 4 Hydrogen undercuts Google's own Carbon and Argon models now, and will cost well under half their price once their December increase lands, so teams on Google should test whether it is good enough for their everyday work before December 1.
arXiv Paper of the Day
AI Isn't Getting Dumber, We're Getting Pickier: Rising Standards in Human Evaluation of Large Language Models
Zihan Chen, Lei Nico Zheng, Di Zhu · arXiv:2611.14333
What it claims: When people say chatbots are "getting worse," part of the reason may be that people's expectations are rising, not that the models changed. The authors studied 1,260,967 blind votes on Chatbot Arena (a public site where people compare two anonymous AI answers) from May 2023 to August 2024, looking only at model versions that never changed.
Key finding: For the very same unchanged models, the share of votes saying "both answers are bad" rose by about 1.7 percentage points a month, and in 14 frequently compared pairs it climbed in every one, from 15.5% to 23.3%, while which model won stayed steady.
Why practitioners should care: User satisfaction scores for an AI feature can fall even when nothing in the product changed, so compare against current alternatives rather than last year's ratings before deciding a model "regressed." The same drift affects the human ratings used to train models and rank leaderboards, so absolute "good or bad" scores age quickly while head-to-head comparisons hold up.
Key finding: For the very same unchanged models, the share of votes saying "both answers are bad" rose by about 1.7 percentage points a month, and in 14 frequently compared pairs it climbed in every one, from 15.5% to 23.3%, while which model won stayed steady.
Why practitioners should care: User satisfaction scores for an AI feature can fall even when nothing in the product changed, so compare against current alternatives rather than last year's ratings before deciding a model "regressed." The same drift affects the human ratings used to train models and rank leaderboards, so absolute "good or bad" scores age quickly while head-to-head comparisons hold up.
×
Click anywhere or press ESC to close
Member discussion