Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
3
million of these cards worldwide in 2027
China's tech giants ordered $35 billion of Nvidia chips that
One Thing to Tell Your Friends
One Thing to Tell Your Friends
Claude ran a 64-hour protein-design project against scorpion venom with almost no human help, and 17 of its designs stuck to the toxins when tested in two real labs.
Summary
TL;DR
Top Stories
Trends
Big tech is building exits from Nvidia, Spending more on AI is not reliably buying better results, and AI agents are getting computers of their own.
Creative AI
Voxtral 3 gives apps a full voice assistant in 50+ languages, Lugh Xtudio lets AI agents edit video beside you on the same timeline, and Simon Willison built a strategy game mostly by asking AI, and it plays multiplayer over sound.
Dev Tools
Research
Business
Surprising
Facebook Pages now pay to post more than about two links a month, 180 AI models now beat the original GPT, and Debate: Is AI killing the joy of programming?.
Worth Watching
GitHub
Leading repos: MadsLorentzen/ai-job (+3,890), cloudflare/security-audit (+1,680), and stablyai/orca (+1,559).
HuggingFace
Leading models: deepseek-ai/DeepSeek-V4.1-Flash (27,737), black-forest-labs/flux-3-image (78,897), and fishaudio/drama (13,495).
Product Hunt
API Pricing
No list-price changes vs. October 19 - the thirteenth straight day of stable prices.
arXiv
One Skill Takes Over: Clashing Agent Skills and a One — A skill's rules were followed 93-99% of the time when loaded alone, but only 61-62% of the time when a second skill for the same job was loaded next to it.
FYI
Hot off the Presses
01
Anthropic's IPO filing shows $16.3 billion in revenue and an $80.7 billion loss
What this means for you: Anthropic, the maker of the Claude chatbot, may soon be a stock you can buy, and its filing shows both how fast AI is growing and how much it costs to build.
Previously: October 14 - Anthropic's quarterly revenue reportedly reached $16.8 billion.
Today: The Wall Street Journal reviewed an updated draft of Anthropic's S-1, the document a company files before an Initial Public Offering (IPO, its first sale of shares to the public). Revenue for January to June 2026 was $16.3 billion, up from about $1.1 billion a year earlier.
The headline loss looks alarming but is mostly paperwork. About $72 billion of it reportedly comes from non-cash charges on convertible notes (loans that can turn into shares), whose value rose as Anthropic's valuation climbed.
“$80.7 billion net loss, about $72 billion of it non-cash”
- About one fifth of revenue came from just two customers - the filing does not name them
- Computing spending rose to nearly $9 billion - from $2.3 billion in the same months of 2025
- Anthropic says it could spend more than $750 billion on cloud contracts - up from about $518 billion in commitments in its September draft
- Its cloud suppliers are also its rivals - TechCrunch says Amazon, Google and Microsoft increasingly treat Anthropic as a competitor, so it is leasing its own data centers
- The listing is reportedly targeted for November - Anthropic declined to comment
02
China's tech giants ordered $35 billion of Nvidia chips that Nvidia reportedly cannot fully supply
What this means for you: The limit on China's AI build-out is shifting from US rules to Nvidia's own factories, which keeps chip supplies tight for everyone.
Reuters reports that ByteDance (TikTok's owner), Alibaba and Tencent ordered more than 3 million Nvidia RTX PRO 5500 cards. The card launched on September 14 and was designed to meet US export rules. Nvidia's top chips are barred from China or need case-by-case licenses.
- Each card costs about $11,400 - so the orders total more than $35 billion at sticker price
- Nvidia reportedly told partners they may get only 5% to 10% of what they asked for
- Nvidia can make only about 3 million of these cards worldwide in 2027 - because it is saving factory space for its next platform, Rubin
- Total Chinese orders reportedly top 6 million cards - with shipments starting in early 2027
- The buyers mainly want them for running AI, not training it - the sources are anonymous and no company commented
03
OpenAI says its new model can rewrite old banking code almost perfectly
What this means for you: The decades-old software behind ATMs and card payments may finally get modernized, which could mean fewer outages and cheaper banking systems.
COBOL is a programming language from 1959 that still runs much of the financial world. OpenAI says COBOL handles more than 40% of online banking and 95% of ATM transactions. The people who know it are retiring, and AI models have had little COBOL to learn from.
In OpenAI's test, a model must convert COBOL programs into modern languages like Java or Python without changing how they behave.
“86.6% for GPT-6.1 Sol vs 5.4% for GPT-5.6 Sol”
- That score used low "reasoning effort" - and cost about 61% as much as the old model
- At high reasoning effort GPT-6.1 Sol reached 97.9% - which OpenAI calls close to the maximum
- Building a database from scratch in 20 languages - GPT-6 Astra passed 87.4% of tests on average, versus 26.8% for the older model
- The scores are OpenAI's own - and real migrations still need expert review and testing
04
DeepSeek gave away the software it uses to run its AI cheaply
What this means for you: Running AI could get cheaper for every company, and Chinese AI no longer depends only on American chips.
Previously: October 19 - the US banned DeepSeek's AI from federal civilian agencies.
Today: DeepSeek, a Chinese AI lab, released DeepInfer for free. It is the inference software (the system that answers user requests) behind DeepSeek's own services. Anyone can use or change it under the Apache 2.0 license.
- It runs on Huawei's Ascend 950 chips as well as Nvidia's newest ones - which helps Chinese companies work around US chip limits
- It currently supports only one model, DeepSeek V4.1 Flash - more models will come as they are released
- Gizmodo puts V4.1 Flash at about $0.035 per million input word pieces (tokens) - versus $10 for OpenAI's GPT-6 Astra on the same service
- The project drew about 12,000 GitHub stars quickly - though key features like a memory cache are still missing
05
Meta's Muse app reached 10 million US downloads in 40 days
What this means for you: ChatGPT's lead among everyday AI apps is no longer safe, and Meta is spending heavily to put its assistant on your phone.
Muse is Meta's AI agent app, launched in September in the US and Canada only. Data firm Sensor Tower says only 11 other apps have reached 10 million US downloads this fast. The only non-games on that list are Disney+ and Meta's own Threads.
- ChatGPT had 3.7 million US downloads in its first 40 days - Claude had 200,000
- US ChatGPT downloads over the last 30 days fell 25% - compared with the 30 days before
- People keep using Muse - 30% are still active after a week, versus 19% for rival AI apps
- Meta put 39% of its October ad spending into Muse - up from 23%
- Muse passed 2.5 million daily users on October 16 - and Meta reports earnings on Wednesday
Trends & Themes
Trends & Themes
Big tech is building exits from Nvidia
Why this matters to you: If AI software can move freely between chip makers, competition could push down the price of the AI services you pay for.
PyTorch, the most popular AI toolkit, grew up tuned for Nvidia. Chip makers are now rewriting that default, and DeepSeek's DeepInfer (see Top Stories) is doing the same from China.
- Google released TorchTPU - letting apps written for Nvidia chips run on Google's own chips with a settings change
- Amazon's TorchNeuron does the same for its Trainium chips - and Amazon says its custom chip business now earns over $25 billion a year
- PyTorch 2.15 runs some AI models over 6 times faster on Apple's Mac chips - than the previous version
- Intel launched a data-center chip that uses cheaper memory - aimed at serving many AI users at once rather than raw speed
Spending more on AI is not reliably buying better results
Why this matters to you: Paying for the most expensive AI setting, or the biggest AI budget, often does not get you better work.
That line is from the SlopCodeBench report's title. The pattern: better instructions and smarter setups are beating brute-force spending.
“xhigh considered harmful”
- One trading firm's AI bill rose from $500,000 a month to $8 million - but across 27 companies, more AI spending often did not ship software faster, per The Pragmatic Engineer
- The top "effort" setting on coding agents cost 2 to 4 times more - and usually scored worse, in the SlopCodeBench test of changing software over time
- IBM made an older, cheaper model top a customer-service leaderboard - by giving it short rules learned from its own past mistakes
- No coding agent fully passed the last stage of most SlopCodeBench problems - only 0 to 3 of 36 per model
AI agents are getting computers of their own
Why this matters to you: If you use AI helpers, they may soon need their own machine, and you will want a way to see what they did on it.
Agents used to live in a chat window. Now they need computers, storage and supervision, which is creating a whole new market.
- Daytona, which rents agents fast throwaway computers, is now worth $1 billion - its revenue rate grew 25-fold from January to September
- Writer Nate Jones says a base $899 Mac mini is too small - for a dedicated AI-agent computer, and urges cleaning up what agents leave running
- Ben Tossell built one personal site to watch his agents work - because he wants to see what they are doing, and readers echoed the wish to avoid "black box magic"
- Professor Ethan Mollick calls the new style "conjuring" - you summon agents and judge results, but lose sight of how they got there
AI labs are drafting the rules they will live under
Why this matters to you: The safety laws that govern the AI you use may be written largely by the companies they regulate.
Anthropic's head of federal affairs says the bill will be revised. Meanwhile OpenAI reportedly plans to pick its preferred path within the next year.
- Anthropic's leaked federal bill would block a model only if it risks 50+ deaths or $1 billion in damage - TechCrunch reports
- Critics say it is weaker than a bipartisan bill already in Congress - and would override state laws like California's
- Sam Altman reportedly told OpenAI staff about three "endgames" for AI - including one where a single company shuts down everyone else's chips
- OpenAI has gone quiet on strict chip export limits to China - Politico reports, a shift that helps Nvidia
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Business & Industry
Business & Industry
Surprising
Surprising & Under-the-Radar
Facebook Pages now pay to post more than about two links a month
Meta ties extra link posts to paid plans starting at $14.99 a month. It drew 399 points on Hacker News. Link posts had already fallen from 9.8% of what people viewed on Facebook in 2022 to about 1.3%. Mastodon: Ashley Gullen on the link limit
180 AI models now beat the original GPT-4, and 19 run on a phone
Ethan Mollick charted 30 months of AI progress. The cost of a correct answer on a graduate-level exam fell from 30 cents to four hundredths of a cent. One Useful Thing: 30 months on the exponential
Debate: Is AI killing the joy of programming?
For: An essay by Jeremy Scheff says coders lost the paid pleasure of solving tricky puzzles. Against: Simon Willison says people who love building working systems find AI "the gift that keeps on giving." Simon Willison: My obituary for programming
Debate: Does AI writing save time, or just move the work to readers?
For saving time: Writers draft much faster. Against: Dima Bolkunov says he reads AI text at least a third slower, and many readers pay that cost for each writer's saving. Dima Bolkunov: On AI writing
Worth Watching
Signals to Track
01
Renting AI chips may be far too cheap
One chip can earn about $40 an hour answering AI requests at full use but rents for $5-6, and that gap cannot last.
Evan Conrad runs SF Compute, a marketplace for AI computing time. He says "one of these numbers has to change." He also warns against betting markets on chip prices before real delivery markets mature. If rental prices rise, the cost of AI apps for ordinary people could rise too.
02
Meta's next open model is built for one gaming graphics card
Muse Glimmer is a 30-billion-parameter (a size measure) agent model meant to run privately on a single consumer Graphics Processing Unit (GPU).
Meta published a technical report on it, according to the AINews newsletter. Small open models that can run tools are improving fast. If this works, capable AI assistants could run at home with no subscription and no data leaving your computer.
03
Fake statistics can now spread through AI answers
One person made 50,000 pages of invented stats for $50, and AI chatbots began citing them.
Within about seven months, the made-up numbers appeared in an academic paper and investment research. The site ranked second only to Statista on tracked AI prompts, even though Google had mostly dropped it. If AI search keeps rewarding freshness over truth, readers will need to check where every AI-quoted number came from.
GitHub Trending
Top Repos Today
The AI job-search kit held the top spot for another day and passed 60,000 stars. Three newcomers joined the list, and all three are add-ons for coding agents: a security auditor from Cloudflare, a "write less code" skill and a tool that maps a whole codebase into a searchable graph.
Rank yesterday: #1 - Holding steady ➡
⭐ Stars today: +3,890 · 📦 Total: 60,433
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 30-45 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 30-45 minutes
What it is: A job-search kit built on Claude Code that judges job postings, tailors your CV, writes cover letters and preps you for interviews on your own computer. Its latest release (version 1.7.3) tightened permissions for its AI helpers, and its release notes say contributors closed 58 bugs in two weeks. Why you'd want it: An AI job-application helper that keeps your data on your machine and is being fixed quickly by a large community.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Needs paid Claude Code plus extra tools |
| Runs locally | Default job boards are Danish |
| Safer permission settings in the latest release | Setup takes longer than most tools here |
GitHub - MadsLorentzen/ai-job-search: The job search that runs on your machine. AI job application framework built on Claude Code: evaluate postings, tailor CVs, write cover letters, prep interviews. Fork it and own it.
The job search that runs on your machine. AI job application framework built on Claude Code: evaluate postings, tailor CVs, write cover letters, prep interviews. Fork it and own it. - MadsLorentzen…
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +1,680 · 📦 Total: 34,064
📜 License: MIT · 👤 By: organization (Cloudflare)
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: organization (Cloudflare)
🎯 Time to value: 15 minutes
What it is: An add-on "skill" that turns a coding agent into a security auditor for your own code. It runs a six-step review in which separate AI helpers hunt for weaknesses and a fresh helper tries to disprove each finding before it reaches the final report. Cloudflare says it is the starting point for its own in-house vulnerability-hunting system. Why you'd want it: A structured, second-opinion security review of your codebase that throws out weak findings instead of flooding you with them.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license, from a major security company | Needs a locked-down sandbox to run safely |
| Every finding is checked by a separate AI helper | Many helper agents means higher usage costs |
| Repeat runs build on earlier ones | Cloudflare says one run found only about half of what repeated runs found |
GitHub - cloudflare/security-audit-skill: A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings - cloudflare/security-audit-skill
GitHub
Rank yesterday: #5 - Rising ↑
⭐ Stars today: +1,559 · 📦 Total: 99,676
📜 License: MIT · 👤 By: organization (Stably AI)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: organization (Stably AI)
🎯 Time to value: 10 minutes
What it is: A desktop and phone app for running several AI coding agents side by side with the subscriptions you already pay for. It shipped two releases today: you can now "rewind" a Claude chat to an earlier message and take it in a new direction, and each chat remembers its own model settings. Why you'd want it: Steer several coding agents at once, even from your phone, and nearly at 100,000 stars.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Several releases a day |
| Uses subscriptions you already pay for | Many agents multiply usage costs |
| New rewind and per-chat settings | A lot of features to learn |
GitHub - stablyai/orca: Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime. - stablyai/orca
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +1,396 · 📦 Total: 169,565
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 5 minutes
What it is: A single set of instructions that makes your AI coding agent behave like a terse senior developer who writes only the code a task needs. Its maker's own test in Claude Code (39 tasks, 5 runs each) reports 53% less code, 41% less time, 26% lower cost and 45% fewer tokens. A recent update adds a bar showing what each task cost and a local page charting your sessions over time. Why you'd want it: Smaller, cheaper changes from your coding agent without giving up tests or safety checks.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, two-command install | Savings figures are the maker's own benchmark |
| Works with 20 coding agents | Terse style may not suit every team |
| Keeps validation and security code by design | Cost bar is Claude Code only |
GitHub - DietrichGebert/ponytail: Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote. - DietrichGebert/ponytail
GitHub
Rank yesterday: #4 - Falling ↓
⭐ Stars today: +1,201 · 📦 Total: 140,722
📜 License: GPL-3.0 · 👤 By: organization (Comfy Org)
🎯 Time to value: 20 minutes
📜 License: GPL-3.0 · 👤 By: organization (Comfy Org)
🎯 Time to value: 20 minutes
What it is: A popular free app for building AI image and video workflows by connecting boxes on a screen, running on your own graphics card. Its newest version (0.42.1) fixes crashes some users hit in the MOSS-Image starter workflow (see Hugging Face below). Why you'd want it: The place where new open image and video models usually show up first.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on your own computer | Strict share-alike license |
| Huge library of add-ons | Some add-ons call paid cloud services |
| Quick fixes for new models | The box-and-wire screen takes learning |
GitHub - Comfy-Org/ComfyUI: The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world. - Comfy-Org/ComfyUI
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +906 · 📦 Total: 130,179
📜 License: Apache-2.0 · 👤 By: organization (Graphify Labs)
🎯 Time to value: 10 minutes
📜 License: Apache-2.0 · 👤 By: organization (Graphify Labs)
🎯 Time to value: 10 minutes
What it is: A tool you call from your AI coding assistant that maps your whole project (code, docs, PDFs, images and video) into a "knowledge graph," a web of linked facts you can question instead of searching file by file. Code is read on your own machine without any AI model, and each link is labeled as either read directly or inferred. Today's release merged 22 community fixes. Why you'd want it: Give your coding agent a map of a large codebase so it reads less and finds connections faster.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Docs and media still need an AI model to process |
| Code mapping stays fully local | Pushes a paid cloud version |
| Works with 20+ assistants | Another file set to keep in sync with your code |
GitHub - Graphify-Labs/graphify: Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store.
Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing…
GitHub
Rank yesterday: #2 - Falling ↓
⭐ Stars today: +1,256 · 📦 Total: 218,868
📜 License: MIT · 👤 By: organization (Anomaly)
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: organization (Anomaly)
🎯 Time to value: 5 minutes
What it is: A free, open-source AI coding assistant that runs in your terminal and works with many AI providers. Its latest 1.x updates keep existing users on the 1.x line, while moving to OpenCode 2 is a separate, opt-in upgrade. Why you'd want it: A capable coding assistant that is not tied to one AI company.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Two version lines to choose between |
| Works with many AI providers | Install steps differ by version |
| Upgrades to version 2 only when you choose | Very fast pace of changes |
GitHub - anomalyco/opencode: The open source coding agent.
The open source coding agent. Contribute to anomalyco/opencode development by creating an account on GitHub.
GitHub
Rank yesterday: #6 - Falling ↓
⭐ Stars today: +555 · 📦 Total: 76,731
📜 License: Apache-2.0 · 👤 By: organization (Headroom Labs)
🎯 Time to value: 5 minutes
📜 License: Apache-2.0 · 👤 By: organization (Headroom Labs)
🎯 Time to value: 5 minutes
What it is: A tool that shrinks logs, files and search results before they reach an AI model, so you pay for fewer tokens. Its makers say coding agents use about 20% fewer tokens with it. Its latest version adds support for Google's Antigravity editor and OpenCode 2. Why you'd want it: Lower AI bills for coding agents with one setup command.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Savings claims are the maker's own |
| Usage reporting is now opt-in | Adds another layer between you and the AI |
| Built-in savings dashboard | Pushes a paid team plan |
GitHub - headroomlabs-ai/headroom: Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server. - headrooml…
GitHub
HuggingFace Trending
Top Models Today
DeepSeek's cheaper, faster update held the top spot for a third day as downloads rose 55%.
📥 Downloads (30d): 27,737 · 📜 License: MIT
👤 By: DeepSeek · 🎯 Task: Image-text-to-text
📐 Size: 763B
👤 By: DeepSeek · 🎯 Task: Image-text-to-text
📐 Size: 763B
What it is: A downloadable AI model that reads images and text, tuned for coding and multi-step agent work. DeepSeek products, including this model, were barred from US federal civilian agencies by a Department of Homeland Security order on October 19. Why you'd want it: A strong, cheap base for coding agents that you are free to host yourself.
| ✓ Pros | ✗ Cons |
|---|---|
| Very permissive license | Needs data-center hardware to run |
| Built for coding and agent tasks | Benchmarks are self-reported |
| Also offered by low-cost hosts | Banned in US federal civilian agencies |
deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Its ComfyUI version rose to second on the trending list with more than 439,000 downloads.
📥 Downloads (30d): 78,897 · 📜 License: FLUX Kommunity License
👤 By: Black Forest Labs · 🎯 Task: Image-to-image
📐 Size: 5.55B
👤 By: Black Forest Labs · 🎯 Task: Image-to-image
📐 Size: 5.55B
What it is: A small image generator and editor that lets you place objects by drawing boxes. Downloads of the main version rose 27% in a day, and most people use it through the ComfyUI app. Why you'd want it: Precise image layouts on a single graphics card.
| ✓ Pros | ✗ Cons |
|---|---|
| Small and fast | License caps revenue for commercial use |
| Box-based placement | Works best with a prompt rewriter |
| Popular in ComfyUI | Gated download |
black-forest-labs/flux-3-image-klein · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Fish Audio's directable voice-acting model grew downloads 14% in a day.
📥 Downloads (30d): 13,495 · 📜 License: Fish Audio research license
👤 By: Fish Audio · 🎯 Task: Text-to-speech
📐 Size: 35.9B (3B active)
👤 By: Fish Audio · 🎯 Task: Text-to-speech
📐 Size: 35.9B (3B active)
What it is: A voice model you can direct like an actor, describing how a line should be read. Only about 3 billion of its parameters work at once, which keeps it efficient for its size. Why you'd want it: Expressive voices for audiobooks, games and videos.
| ✓ Pros | ✗ Cons |
|---|---|
| Directable performances | Research license limits commercial use |
| Efficient design | Voice cloning raises consent issues |
| Free to download without approval | Large to download |
fishaudio/drama-3 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Meta's biggest downloadable Muse model climbed to fourth on the trending list as downloads rose 23%.
📥 Downloads (30d): 2,592 · 📜 License: Muse Model License
👤 By: Meta · 🎯 Task: Image-text-to-text
📐 Size: 822B (42B active)
👤 By: Meta · 🎯 Task: Image-text-to-text
📐 Size: 822B (42B active)
What it is: Meta's first big open-weight Muse model, which reads text, images, video and audio. Downloads need manual approval, and Groq also offers it as a preview for testing. Why you'd want it: A top American model you can download and run yourself.
| ✓ Pros | ✗ Cons |
|---|---|
| Strong US open-weight model | Needs very large hardware |
| Handles many media types | Custom license with conditions |
| Hosted preview available on Groq | Manual approval to download |
meta-models/Muse-Spark-Open-816B-A42B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Google's small multimedia search model held fifth as download growth stayed flat at 3%.
📥 Downloads (30d): 172,474 · 📜 License: Apache-2.0
👤 By: Google · 🎯 Task: Feature extraction (embeddings)
📐 Size: 0.74B
👤 By: Google · 🎯 Task: Feature extraction (embeddings)
📐 Size: 0.74B
What it is: A small model that turns text, images, video and audio into number lists (embeddings) that can be searched together. It is small enough for a laptop. Why you'd want it: One search index for documents, screenshots, video and audio.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Text gains over version 1 are modest |
| One index for every media type | Google's own benchmark numbers |
| Runs on-device | Tool support still thin |
google/embeddinggemma-2 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Downloads rose 36% in a day, and ComfyUI's own packaged copy has passed 924,000 downloads.
📥 Downloads (30d): 4,547 · 📜 License: Apache-2.0
👤 By: OpenMOSS · 🎯 Task: Text-to-image
📐 Size: 5.08B
👤 By: OpenMOSS · 🎯 Task: Text-to-image
📐 Size: 5.08B
What it is: A small image generator that lets you describe where each object and caption goes. It can write English and Chinese text inside images. Why you'd want it: Posters and diagrams with text in the right place, with commercial use allowed.
| ✓ Pros | ✗ Cons |
|---|---|
| Commercial use allowed | Slow, 50 steps per image |
| Layout control for text | Early ComfyUI template crashes needed a fix |
| Fits on one graphics card | No independent benchmark scores |
OpenMOSS-Team/MOSS-Image-SFT · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Lightricks' open video-and-sound model is ninth on the trending list with more than 2 million downloads.
📥 Downloads (30d): 2,039,447 · 📜 License: LTX-2.x Community License
👤 By: Lightricks · 🎯 Task: Image-to-video
📐 Size: about 22B (per its file names)
👤 By: Lightricks · 🎯 Task: Image-to-video
📐 Size: about 22B (per its file names)
What it is: A model that creates video with matching sound from text, images or other video, and can keep characters consistent across several shots. It can render up to 4K (3840 by 2176 pixels) and has options for running on smaller graphics cards. Why you'd want it: Open video generation with built-in audio that you can run and fine-tune yourself.
| ✓ Pros | ✗ Cons |
|---|---|
| Video and sound in one pass | Companies over $10M annual revenue need a paid license |
| Free commercial use for smaller companies | Download requires agreeing to marketing terms |
| Low-memory options available | Needs a recent NVIDIA setup |
Lightricks/LTX-2.5 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Liquid AI's largest small-footprint model, built to run fast, private agents on your own device.
📥 Downloads (30d): 15,410 · 📜 License: LFM Open License 1.0
👤 By: Liquid AI · 🎯 Task: Text generation
📐 Size: 24B (about 2.5B active)
👤 By: Liquid AI · 🎯 Task: Text generation
📐 Size: 24B (about 2.5B active)
What it is: A text-only reasoning model where only about 2.5 billion of its 24 billion parameters work at once, so it runs quickly on laptops. Liquid AI reports 184 tokens per second on an Apple M5 Max and a score of 48.09 on the LiveClaw agent test, versus 38.19 for the larger Qwen3.6-35B-A3B. Why you'd want it: A private, on-device assistant for tool-using tasks that does not need a cloud connection.
| ✓ Pros | ✗ Cons |
|---|---|
| Fast on ordinary hardware | License has a $10M annual revenue threshold |
| Ready-made versions for laptops and phones | Benchmarks are the maker's own |
| 131,000-token memory, 16 languages | Text only, no images |
LiquidAI/LFM2.5-24B-A2B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Product Hunt
AI Launches Today
The easiest way for AI agents to pay for their own tools
🔥 Upvotes: 251 · 👤 By: Composio
💰 Pricing: free to try, pay as you go · 🏷 Category: Developer tools
💰 Pricing: free to try, pay as you go · 🏷 Category: Developer tools
Normally, giving an AI agent a new tool means signing up, getting an access key and adding billing for each service. Composio Instant gives your agent one wallet so it can use web search, contact lookup, voiceovers, transcription and more on demand, with no separate sign-ups. It builds on Composio's library of 1,000+ ready-made agent tools. Verdict: A real time-saver for agent builders juggling many services, but set spending limits before handing an agent a wallet.
Composio: Your agent acts. We'll handle the rest. | Product Hunt
Your AI agent knows what to do, Composio gives it the tools to actually do it. 1,000+ production-ready tools with managed auth, smart tool resolution, and sandboxed execution. Your agent calls the tool, we handle OAuth, permissions, and reliability. Five lines of code to go from chatbot to general-purpose agent. Used by teams at Zoom, Glean, and HubSpot.
Product Hunt
Smarter search that gives AI agent richer data at lower cost
🔥 Upvotes: 138 · 👤 By: Exa
💰 Pricing: free options; Exa Lite $4 and Exa Auto $7 per 1,000 searches · 🏷 Category: Search for AI agents
💰 Pricing: free options; Exa Lite $4 and Exa Auto $7 per 1,000 searches · 🏷 Category: Search for AI agents
Exa's search service for AI agents gets its biggest quality update, adding billions of new pages, including 1.27 billion online store listings and 211 million+ local businesses. A new "Dynamic Highlights" feature returns only the relevant parts of a page, which Exa says uses 95% fewer tokens than the full text. Verdict: A strong pick for agents that do research or shopping lookups, with clear per-search prices that make costs easy to predict.
exa.ai: Search API for AI applications | Product Hunt
Find perfect lists of leads, candidates or anything in the web. Websets’ semantic search lets you ask in plain English, it agentically sources results and enriches them with you ask for - like emails, tags, and more - all in the time it takes to brew a coffee.
Product Hunt
An always-on agent that keeps your docs up to date
🔥 Upvotes: 122 · 👤 By: Mintlify
💰 Pricing: free option listed; Autopilot plan pricing not stated · 🏷 Category: Documentation
💰 Pricing: free option listed; Autopilot plan pricing not stated · 🏷 Category: Documentation
An agent that watches your code, Slack, project tracker and support tickets, then drafts documentation updates when something changes. Nothing publishes on its own: every change arrives as a pull request (a proposed edit) for your team to review. Mintlify's own survey found only 24% of teams update docs the same day a product changes. Verdict: A practical fix for docs that fall behind fast-moving AI-written code, as long as someone actually reviews the drafts.
Mintlify: The intelligent knowledge platform | Product Hunt
The next generation of documentation. AI-native, beautiful out-of-the-box, and built for collaboration.
Product Hunt
Open models, built for agents.
🔥 Upvotes: 110 · 👤 By: Jaber Jaber and Osama Jaber
💰 Pricing: free option; coding plans from $10/month · 🏷 Category: Developer tools
💰 Pricing: free option; coding plans from $10/month · 🏷 Category: Developer tools
A flat monthly plan that plugs coding agents like Claude Code, Codex and OpenCode into fast open models such as GLM 5.3, DeepSeek V4.1 Flash and Qwen3.8 27B. It says smart routing reuses cached work up to 99.5% of the time and that it keeps no copies of your data. Verdict: A cheap way to try open models inside your usual coding agent, though one early commenter questioned whether the entry plan's limits are generous enough.
RunInfra: Fast open models for your coding agent | Product Hunt
Plug Claude Code, Codex, OpenCode, Cline, Droid or Grok Build into fast open models like GLM 5.3, DeepSeek V4.1 Flash and Qwen3.8 27B. Plans from $10 a month, up to 99.5% cache hits so your plan goes further, and zero data retention.
Product Hunt
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| 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 Luna | $0.10 (up to 272K; $0.20 above) | $0.50 ($0.75 above 272K) | 1.05M 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 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 |
| Groq | Muse Spark Open (preview) | Contact sales | Contact sales | 131K tokens |
No list-price changes vs. October 19 - the thirteenth straight day of stable prices. Groq labels Qwen3.8-27B and Muse Spark Open as preview models meant for testing, not production.
What this means: Gemini 4 Argon's introductory price has 41 days left before it doubles. With flagship prices frozen, the action has moved to cutting how many tokens you use: tools like Ponytail and Headroom (see GitHub above) target the same bill.
What this means: Gemini 4 Argon's introductory price has 41 days left before it doubles. With flagship prices frozen, the action has moved to cutting how many tokens you use: tools like Ponytail and Headroom (see GitHub above) target the same bill.
arXiv Paper of the Day
One Skill Takes Over: Clashing Agent Skills and a One-File Fix
Parsa Razmara, Seyedarmin Azizi, Ardalan Aryashad · arXiv:2610.25064
What it claims: Agent "skills" are instruction files that teach a coding agent how to do one job, and people often install two skills from different authors for the same job. The authors built a test set of 16 such skill pairs and found that, across seven models from five families, the AI usually follows one skill completely and silently drops the other, driven mostly by the worked examples inside each file.
Key finding: A skill's rules were followed 93-99% of the time when loaded alone, but only 61-62% of the time when a second skill for the same job was loaded next to it.
Why practitioners should care: If you install skills or rules files from several sources, some of your rules are probably being ignored without warning. The paper's fix is simple: merge overlapping skills into one file with a single correct worked example, which brought rule-following back to 96-99% of what was achievable on all seven models.
Key finding: A skill's rules were followed 93-99% of the time when loaded alone, but only 61-62% of the time when a second skill for the same job was loaded next to it.
Why practitioners should care: If you install skills or rules files from several sources, some of your rules are probably being ignored without warning. The paper's fix is simple: merge overlapping skills into one file with a single correct worked example, which brought rule-following back to 96-99% of what was achievable on all seven models.
×
Click anywhere or press ESC to close
Member discussion