Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
27,
iPadOS 27, macOS 27, tvOS 27, watchOS
Tomorrow's WWDC Will Reshape How 2 Billion Apple Devices Use
4.5
and later versions with MCP connections (a
A Senior Engineer Says AI Has Eroded Every Pillar of Their C
73%
of engineering teams now use AI coding
AI Is Rewriting the Designer-Developer Handoff
One Thing to Tell Your Friends
One Thing to Tell Your Friends
Apple is about to let you choose whether Siri thinks with Google's brain, OpenAI's brain, or Anthropic's brain - and tomorrow's keynote is Tim Cook's last.
Summary
TL;DR
Top Stories
Trends
AI Is Rewriting the Designer, The Software Engineering Career Crisis Is Going Mainstream, and Government AI Ownership Is No Longer Fringe.
Creative AI
Google Veo 3.1 Leads Video Generation, But Competition Is Closing In.
Dev Tools
Lathe: An LLM Tool That Teaches You Instead of Thinking for You, Grok Build: xAI's Terminal Coding Agent Enters Beta, and Claude Desktop for Linux: 424 HN Points and Counting.
Research
Business
Uber's AI Budget Crisis Has a New Chapter: The COO Questions ROI, xAI Wins First Federal Contract: Grok 4 for Government, and OpenAI Models Now Available on AWS.
Surprising
A 10, xAI Got a Government Contract for $0.42 Per Agency, and The "Being Good at AI" Bar Is Embarrassingly Low.
Worth Watching
Apple's AI Provider Choice Could Commoditize Model Companies, Colorado's AI Act Takes Effect June 30 - First US Risk, and The "Specialist to Generalist" Economic Trap May Define the Next Decade.
GitHub
Leading repos: RyanCodrai/turbovec (+1,533), NousResearch/hermes (+1,117), and mvanhorn/last30days (+1,097).
HuggingFace
Leading models: nvidia/LocateAnything (116k), google/gemma-4-12B (435k), and unsloth/gemma-4-12b-it (568k).
API Pricing
What this means:** Anthropic and OpenAI are now price-matched at $5/M input for their flagship models (Opus 4.8 vs GPT-5.5), but Anthropic offers a 1M context window at flat rates while OpenAI charges 2x input and 1.5x output for prompts exceeding 272K tokens.
arXiv
Beyond Scaling — Hierarchical, graph-based, and swarm-style agent organization assumptions all change fundamentally when agents run on edge devices with limited compute, requiring new coordination patterns that do not exist in cloud-scale systems.
FYI
Hot off the Presses
01
Tomorrow's WWDC Will Reshape How 2 Billion Apple Devices Use AI
What this means for you: Starting this fall, your iPhone will let you pick which AI company powers Siri - Google, OpenAI, or Anthropic. If you have never used an AI assistant that actually works, Apple is betting this is the version that changes your mind.
Apple's Worldwide Developers Conference keynote starts June 8 at 10 a.m. Pacific. It is expected to be Tim Cook's final keynote before John Ternus becomes CEO on September 1.
The multi-model approach is unprecedented for a major platform. Rather than locking users into one AI provider, Apple is positioning itself as the distribution layer - a strategy that could reshape how AI companies compete for consumers.
""$1 billion per year for Google's AI brain inside every iPhone - the largest single AI licensing deal ever reported.""
- Siri is being completely rebuilt on a custom 1.2-trillion-parameter Google Gemini model that Apple is licensing for approximately $1 billion per year
- Multi-model selection - a new Extensions system will let users choose whether ChatGPT, Google Gemini, or Anthropic's Claude handles Apple Intelligence features
- New capabilities include multi-step task handling, conversational context awareness, a dedicated Siri app, Dynamic Island integration, and camera-based features like reading nutrition labels
- New operating systems - iOS 27, iPadOS 27, macOS 27, tvOS 27, watchOS 27, and visionOS 27 will all be announced
02
A Senior Engineer Says AI Has Eroded Every Pillar of Their Career
What this means for you: If you work in software engineering or are considering it as a career, this essay lays out the case that three types of expertise once considered safe from automation - deep domain knowledge, debugging skill, and architectural taste - are all losing their economic value at the same time.
A finance/payments engineer with 10 years of experience wrote a post that became the most-discussed item on Hacker News today, with 745 points and 717 comments.
The author notes that brilliant former colleagues remain unemployed despite deep specialization. Career pivots toward frontier ML research feel impractical for most people.
- Domain knowledge is now "promptable" - specialized expertise in PCI compliance (the security standard for credit card processing), double-entry ledgers, and payment system lifecycles can be synthesized from public documentation by LLMs (Large Language Models - the AI systems behind tools like ChatGPT and Claude)
- Debugging hit hardest - Claude 4.5 and later versions with MCP connections (a protocol that lets AI tools interact with external systems) achieve approximately 90% one-shot bug resolution on race conditions and corner-cases that previously required 1-2 days of manual investigation
- Architecture taste is devaluing - organizations are accepting lower-quality codebases because AI-generated code is fast and cheap, even if the design would not pass traditional review
- The economic trap - when AI turns every specialist into a generalist, the market price of a generalist falls. "Domain expertise no longer provides competitive advantage when it is now promptable."
03
A Jane Street Designer Now Uses Claude More Than Figma
What this means for you: The traditional pipeline of design mockup, developer handoff, and review cycles may be collapsing. One designer at a major trading firm builds production-ready features directly with AI, skipping the design tool entirely.
Edwin Morris, a designer at Jane Street (one of the world's largest quantitative trading firms), published a detailed account of how Claude Code has replaced his design workflow.
The post generated 235 points and 214 comments on Hacker News, with debate centering on whether this represents efficiency gains or a loss of design rigor.
- Old process: write problem descriptions, build Figma mockups, write proposals, review implementation with developers
- New process: open an editor, build server, and Claude with a problem description as the prompt - then iterate until the feature is done
- Scale: Morris now ships features with 2,000+ line code diffs built entirely through Claude, not just small UX tweaks
- Speed: refinements to a SQL input tool - submit button, keyboard shortcuts, copy adjustments, AI prompt tuning - would have taken "days or weeks of engineering and design back-and-forth" at previous roles
- The key insight: "Claude gave me free, unlimited iteration, unbothered when I changed my mind for the 50th time"
- Acknowledged tradeoff: reviewers now receive fully-baked features rather than design proposals for collaborative feedback, which may constrain creative exploration
04
Trump and Sanders Both Want the Government to Own Equity in AI Companies
What this means for you: Two politicians who agree on almost nothing both believe the government should own a piece of the companies building the most powerful AI systems. If either proposal advances, it would be the first time the US government has taken equity stakes in private technology companies.
President Trump and Senator Bernie Sanders publicly endorsed government equity stakes in AI companies this week, creating unexpected political alignment on AI governance.
This is a story with no precedent in US technology policy. The government has subsidized, regulated, and broken up tech companies - but it has never demanded ownership stakes.
- Trump called the concept "a beautiful thing" after Sam Altman privately pitched the idea to administration officials starting in early 2025
- Sanders proposed the American AI Sovereign Wealth Fund Act - a 50% one-time stock tax paid in equity, creating a federal sovereign wealth fund with board voting rights
- The bill targets frontier AI companies - specifically OpenAI, Anthropic, and xAI - while notably excluding Google and Meta
- The divergence: Trump appears open to a negotiated, voluntary equity arrangement; Sanders wants mandatory transfer backed by legislation
- Context: the combined private valuations of the three targeted companies now exceed $2.5 trillion
05
An Entire Office Document Viewer Was Built by Claude
What this means for you: A developer used Claude to write every line of a project that renders Word, Excel, and PowerPoint files in a web browser - including the Rust code, TypeScript code, tests, and tooling. If AI can build something this complex through conversation alone, the definition of "solo developer" is changing.
The Silurus/ooxml project renders Microsoft Office files (.docx, .xlsx, .pptx) directly to an HTML Canvas element, and its creator states the entire codebase was implemented by Claude through iterative prompting.
The project reached 107 points on Hacker News. It represents one of the most complex publicly documented examples of AI-generated software - not a prototype or demo, but a full-featured document renderer with cross-platform support.
- Architecture: Rust parsers compile to WebAssembly (a technology that lets code written in languages like Rust run safely in web browsers), with Web Workers handling parsing and a Canvas 2D renderer drawing the output
- DOCX support: pages, headers, tables, text styling, math equations, images, track changes, and footnotes
- XLSX support: multiple sheets, merged cells, frozen panes, conditional formatting, charts, sparklines, and zoom controls
- PPTX support: master/layout inheritance, 130+ preset shapes, gradients, pattern fills, tables, charts, video playback, and vertical text
- Framework bindings for React, Vue, Angular, Svelte, SolidJS, and Qwik
- Companion tools include a VS Code extension, an MCP server for AI agents, and a markdown export CLI
Trends & Themes
Trends & Themes
AI Is Rewriting the Designer-Developer Handoff
Why this matters to you: The traditional pipeline of mockup, spec, handoff, and review may be the next workflow AI collapses - and if you work in either design or development, your daily process could look very different within a year.
The pattern is clear: when iteration becomes free, the separation between "designing it" and "building it" stops making economic sense. The people who thrive will be those who can do both - or who redefine what design means in this context.
- Jane Street's designer builds production features with 2,000+ line diffs directly in Claude, skipping Figma for entire applications
- Figma itself launched a Claude Code integration that converts production UIs into editable Figma frames, creating a bidirectional loop between code and design
- 73% of engineering teams now use AI coding tools daily according to Pragmatic Engineer's survey of 15,000 developers, blurring the line between who designs and who implements
- Claude Code and Cursor now each claim 18% developer adoption at work, tying for second place behind GitHub Copilot at 29%
The Software Engineering Career Crisis Is Going Mainstream
Why this matters to you: The question is no longer whether AI changes software engineering jobs, but how fast the economic value of traditional programming skills is declining - and what replaces them.
Two reactions are emerging simultaneously: existential anxiety about career relevance, and new tools designed to preserve human learning in an AI-accelerated world. Both are rational responses.
- A senior engineer's "LLMs are eroding my career" essay hit 745 points on HN, the day's most-discussed post, articulating what many feel but few have said publicly
- Claude achieves ~90% one-shot bug resolution on distributed system race conditions according to the author - tasks that previously took 1-2 days
- Uber's 95% monthly AI tool adoption among engineers shows the shift is not optional; it is organizational policy
- The "specialist to generalist" trap - when domain expertise becomes promptable, every specialist competes as a generalist, and generalist wages fall
- Counterpoint: Lathe (214 HN points today) explicitly pushes back, building a tool that uses LLMs to teach domain knowledge rather than skip past it
Government AI Ownership Is No Longer Fringe
Why this matters to you: When both the sitting president and his most prominent progressive critic agree that the government should own pieces of AI companies, the policy window has shifted from "if" to "how much."
The US has never taken equity in private tech companies. Both proposals face massive legal and practical hurdles. But the bipartisan consensus that something should happen makes some form of government AI ownership more likely than it was a week ago.
- Trump endorsed the concept after private discussions with OpenAI's Sam Altman dating back to early 2025
- Sanders proposed 50% mandatory equity transfer via the American AI Sovereign Wealth Fund Act, targeting OpenAI, Anthropic, and xAI
- xAI already secured a federal contract - the GSA signed a OneGov agreement making Grok 4 available government-wide at $0.42 per agency for 18 months
- Combined private valuations of targeted companies exceed $2.5 trillion, making the stakes enormous in dollar terms
Apple's Multi-Model Strategy Could Reshape AI Competition
Why this matters to you: If Apple lets you choose your AI provider the way you choose your default browser, the entire competitive landscape shifts from "which AI is best" to "which AI works best for you specifically."
If Apple becomes the referee rather than a player, it solves a consumer problem (choice) while creating a business problem for AI companies (commoditization). The company that wins the default slot on 2 billion devices has an enormous advantage - and paying $1 billion/year for it may be cheap.
- Apple's Extensions system will let users swap between ChatGPT, Gemini, and Claude for Apple Intelligence features
- Google's $1 billion/year licensing deal for the default Gemini-powered Siri sets a new benchmark for AI distribution costs
- Apple controls distribution to 2+ billion active devices - becoming the AI layer means reaching more users than any model provider can alone
- The precedent: Apple's browser choice screen in the EU already showed that default selection dramatically shapes market share
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Edge AI Agents: Running Intelligence on Your Device, Not in the Cloud
What this means for you: Researchers are working on running AI agents directly on phones and laptops instead of relying on cloud servers - which would make them faster, cheaper, and private by default.
- "Beyond Scaling: Agents Are Heading to the Edge" (arXiv) argues that on-device deployment is technically plausible through small models, quantization (a technique that shrinks AI models to run on cheaper hardware), and mobile-optimized inference
- Hierarchical agent organization changes fundamentally at the edge - the paper explores how swarm-style coordination works when individual agents have limited compute
- Practical implication: AI assistants that work without an internet connection and never send your data to a server
JetBrains Mellum2: A Coding Model That Shows Its Reasoning
What this means for you: JetBrains (the company behind IntelliJ and other popular code editors) released a 12B-parameter coding model that explains its reasoning as it works - a "thinking" model for code.
- 12B parameters with only 2.5B active (a mixture-of-experts design that keeps the model fast)
- 16,900 downloads in 6 days on Hugging Face
- Designed to integrate with JetBrains IDEs for code completion and explanation
Business & Industry
Business & Industry
Surprising
Surprising & Under-the-Radar
A 10-Year Debugging Veteran Says Claude Resolves 90% of Bugs in One Shot
The "LLMs eroding my career" author's specific claim - that Claude 4.5+ with MCP connections achieves ~90% one-shot resolution on distributed system race conditions that previously took 1-2 days - is one of the most concrete performance benchmarks a practitioner has publicly shared. If true across domains, it redefines what "debugging skill" means as a career asset.
xAI Got a Government Contract for $0.42 Per Agency
The GSA's OneGov deal priced Grok 4 access at less than the cost of a candy bar per federal agency. The pricing suggests xAI is prioritizing government distribution over revenue - a land grab for institutional adoption that could lock competitors out of the fastest-growing enterprise segment.
The "Being Good at AI" Bar Is Embarrassingly Low
Ruben Hassid's viral essay argues that the single most effective AI technique - asking Claude to ask you questions before doing work - puts users "in the top 99.9% of the population." If the skill ceiling is genuinely that low, the bottleneck to AI adoption is not capability or cost but willingness to try.
A Failed Hackathon Project Is More Honest Than Most AI Demos
The "Amazing Digital Dentures" write-up on Hugging Face documents a project that failed to get Nemotron 30B to generate working Three.js games, pivoting to simple HTML toys instead. In a landscape saturated with cherry-picked demos, an honest account of model limitations is more useful than another success story.
Worth Watching
Signals to Track
01
Apple's AI Provider Choice Could Commoditize Model Companies
Apple is about to become the world's largest AI distribution platform - and the companies fighting for default status may find that winning costs more than losing.
If Apple's Extensions system lets users freely swap between Gemini, ChatGPT, and Claude, the AI providers compete on a level playing field inside someone else's ecosystem. Google is paying $1 billion/year for the default slot. The question is whether being the default on 2 billion devices is worth that price - or whether it traps you in a race to outbid competitors indefinitely. Watch for the licensing terms when WWDC details drop tomorrow.
02
Colorado's AI Act Takes Effect June 30 - First US Risk-Based AI Law
The first comprehensive US state law regulating AI in employment decisions goes live in three weeks, and most companies are not ready.
Colorado's Artificial Intelligence Act requires companies using "high-risk AI systems" to run impact assessments, notify workers before AI-based employment decisions, provide appeal mechanisms, and publish statements about AI systems in use. It takes effect June 30, 2026, making it the earliest enforceable state-level AI regulation in the US. Companies using AI for hiring, promotion, or workforce decisions should be preparing compliance plans now.
03
The "Specialist to Generalist" Economic Trap May Define the Next Decade
When AI makes everyone a generalist, the economics of specialization break - and the career strategies that worked for the last 20 years stop working.
Today's most-discussed HN post articulates something economists have theorized but few practitioners have felt: that domain expertise loses its premium when the knowledge is "promptable." The counterargument - that taste, judgment, and system-level thinking remain human advantages - has not yet been tested at the scale AI is now reaching. Worth watching whether the job market data supports the theory in the next 12 months.
04
AI-Built Software Is Getting Genuinely Complex
An entire Office document viewer - Rust parsers, WebAssembly, Canvas rendering, 130+ shapes, six framework bindings - was built by Claude through conversation. The ceiling of what AI can build solo keeps rising.
The Silurus/ooxml project is notable not because AI wrote code, but because of the scope: cross-language compilation, format-specific parsers, a rendering pipeline, and framework integrations that would take a small team months. If this level of complexity is achievable through iterative prompting, the definition of "solo developer" is expanding to include projects that would have required a team.
GitHub Trending
Top Repos Today
Rank yesterday: New entry 🆕
⭐ Stars today: +1,533 · 📦 Total: 7,072
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 10 minutes
What it is: A Rust-based vector search index implementing Google Research's TurboQuant algorithm with Python bindings. It compresses vector embeddings to 16x smaller size while maintaining search speed and recall quality. A 10-million-document corpus fits in 4 GB instead of 31 GB with standard float32 storage. Why you'd want it: If you are building a RAG (Retrieval-Augmented Generation) application or any search system using vector embeddings, this lets you handle much larger datasets on much cheaper hardware. Integrates with LangChain, LlamaIndex, and Haystack.
| ✓ Pros | ✗ Cons |
|---|---|
| 16x memory reduction vs float32 | New project, limited production battle-testing |
| Faster than FAISS on benchmarks | Rust dependency adds build complexity |
| Online ingest without training phases | Community and documentation still growing |
GitHub - RyanCodrai/turbovec: A vector index built on TurboQuant, written in Rust with Python bindings
A vector index built on TurboQuant, written in Rust with Python bindings - RyanCodrai/turbovec
GitHub
Rank yesterday: #1 on June 5 - Falling ↓
⭐ Stars today: +1,117 · 📦 Total: 185,860
📜 License: Apache 2.0 · 👤 By: Organization (Nous Research)
🎯 Time to value: 15 minutes
📜 License: Apache 2.0 · 👤 By: Organization (Nous Research)
🎯 Time to value: 15 minutes
What it is: A general-purpose AI agent framework that "grows with you" - adapting its capabilities and context over time. Built by Nous Research, the open-source AI research lab known for fine-tuned Hermes models. Why you'd want it: An open-source agent that learns your preferences and workflows rather than starting fresh each session.
| ✓ Pros | ✗ Cons |
|---|---|
| Massive community (185k+ stars) | Can be overwhelming for simple use cases |
| Active development and model updates | Requires understanding of agent concepts |
| Adapts to user over time | Resource-heavy for full deployment |
GitHub - NousResearch/hermes-agent: The agent that grows with you
The agent that grows with you. Contribute to NousResearch/hermes-agent development by creating an account on GitHub.
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +1,097 · 📦 Total: 30,772
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 5 minutes
What it is: An AI agent research skill that searches Reddit, X, YouTube, TikTok, Hacker News, and Polymarket simultaneously, scoring results by real engagement metrics (upvotes, likes, prediction market odds) rather than search engine rankings. Synthesizes findings into grounded summaries with citations. Why you'd want it: Instead of manually checking six platforms to understand what people are actually saying about a topic, this searches them all at once and ranks by genuine community engagement rather than SEO.
| ✓ Pros | ✗ Cons |
|---|---|
| Multi-platform search in one query | Requires Claude Code or compatible agent |
| Engagement-scored rather than SEO-ranked | Platform Application Programming Interface (API) rate limits may apply |
| Covers prediction markets for forward-looking data | 30-day window limits historical research |
GitHub - mvanhorn/last30days-skill: AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary
AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary - mvanhorn/last30days-skill
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +1,104 · 📦 Total: 36,526
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 5 minutes
What it is: Calls itself "The Anti-Slop Frontend Framework for AI Agents." It provides design instruction sets (SKILL.md files) that integrate with Claude, Cursor, and Codex to prevent AI-generated interfaces from looking generic. Includes adjustable design "dials" and image-generation utilities for creating reference designs before implementation. Why you'd want it: If you are tired of every AI-built UI looking the same - same spacing, same typography, same blandness - this gives your AI agent design taste.
| ✓ Pros | ✗ Cons |
|---|---|
| Directly addresses the "AI slop" problem | Design quality is subjective |
| Multiple visual style presets | Requires Claude Code or compatible tool |
| Image generation for design references | Additional prompting overhead |
GitHub - Leonxlnx/taste-skill: Taste-Skill - gives your AI good taste. stops the AI from generating boring, generic slop
Taste-Skill - gives your AI good taste. stops the AI from generating boring, generic slop - Leonxlnx/taste-skill
GitHub
Rank yesterday: #5 - Holding steady ➡
⭐ Stars today: +555 · 📦 Total: 27,200
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 10 minutes
What it is: An open-source implementation of Google's NotebookLM with more flexibility and features. Lets you upload documents and have AI-generated conversations, summaries, and audio overviews from your own content. Why you'd want it: NotebookLM but self-hosted, with more control over models and output formats.
| ✓ Pros | ✗ Cons |
|---|---|
| Self-hosted, full data control | Requires more setup than hosted NotebookLM |
| Model flexibility (use any provider) | Audio generation quality varies by model |
| Active community (27k+ stars) | Missing some NotebookLM-exclusive features |
GitHub - lfnovo/open-notebook: An Open Source implementation of Notebook LM with more flexibility and features
An Open Source implementation of Notebook LM with more flexibility and features - lfnovo/open-notebook
GitHub
Rank yesterday: Not in top 10 - Rising ↑
⭐ Stars today: +338 · 📦 Total: 47,458
📜 License: Apache 2.0 · 👤 By: Organization (Agentic AI Foundation / Linux Foundation)
🎯 Time to value: 10 minutes
📜 License: Apache 2.0 · 👤 By: Organization (Agentic AI Foundation / Linux Foundation)
🎯 Time to value: 10 minutes
What it is: A general-purpose AI agent available as a desktop app, CLI, and API. Compatible with 15+ LLM providers and 70+ extensions via Model Context Protocol. Handles code, workflows, research, writing, automation, and data analysis. Why you'd want it: An open-source, provider-agnostic AI agent that works with whatever model you prefer and extends through MCP.
| ✓ Pros | ✗ Cons |
|---|---|
| 15+ LLM provider support | Can be complex to configure |
| 70+ MCP extensions | Heavier than focused single-purpose tools |
| Linux Foundation backing | Desktop app still maturing |
GitHub - aaif-goose/goose: an open source, extensible AI agent that goes beyond code suggestions - install, execute, edit, and test with any LLM
an open source, extensible AI agent that goes beyond code suggestions - install, execute, edit, and test with any LLM - aaif-goose/goose
GitHubRank yesterday: Not in top 10 - Rising ↑
⭐ Stars today: +197 · 📦 Total: 115,301
📜 License: MIT · 👤 By: Organization (GGML)
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: Organization (GGML)
🎯 Time to value: 15 minutes
What it is: The foundational C/C++ library for running large language models locally on consumer hardware. Supports quantization, Graphics Processing Unit (GPU) acceleration, and dozens of model architectures. Why you'd want it: If you want to run AI models on your own computer without sending data to the cloud, this is the most mature and widely-used tool for doing so.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on consumer hardware including laptops | Requires command-line comfort |
| Supports nearly all open model formats | Performance varies by hardware |
| 115k stars, extremely mature | Configuration can be intimidating |
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
HuggingFace Trending
Top Models Today
A vision model that finds any object in any image when given a text description - point to "the red mug behind the laptop" and it draws a precise boundary around it.
📥 Downloads (30d): 116k · 📜 License: Apache 2.0
👤 By: NVIDIA · 🎯 Task: Image-Text-to-Text
📐 Size: 4B
👤 By: NVIDIA · 🎯 Task: Image-Text-to-Text
📐 Size: 4B
What it is: A 4-billion-parameter model that takes an image and a text query and returns precise segmentation masks around the described objects. Think "Ctrl+F for images." Why you'd want it: If you need to automatically find and isolate objects in photos for editing, inventory management, accessibility, or data labeling without training a custom model.
| ✓ Pros | ✗ Cons |
|---|---|
| Works with natural language queries | 4B parameters requires decent GPU |
| Apache 2.0 license for commercial use | Segmentation quality varies with ambiguous queries |
| Strong community adoption (1.5k likes) | Best results require thoughtful prompting |
nvidia/LocateAnything-3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Google's newest open-weight model handles text, images, and audio in one architecture - a true multimodal model anyone can download.
📥 Downloads (30d): 435k · 📜 License: Gemma
👤 By: Google · 🎯 Task: Any-to-Any
📐 Size: 12B
👤 By: Google · 🎯 Task: Any-to-Any
📐 Size: 12B
What it is: A 12-billion-parameter instruction-tuned model from Google that processes text, images, and audio inputs and generates text outputs. Part of the Gemma 4 family released alongside quantized versions for mobile deployment. Why you'd want it: A capable multimodal model you can run locally, fine-tune, and deploy commercially - the open-weight equivalent of proprietary models that cost per-token to use.
| ✓ Pros | ✗ Cons |
|---|---|
| True multimodal (text + image + audio) | Gemma license has some restrictions |
| Strong benchmark performance for size | 12B still needs a decent GPU |
| Quantized versions available for mobile | Not as capable as 70B+ models |
google/gemma-4-12B-it · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
The same Google Gemma 4 model, pre-packaged in the format that lets you run it on a laptop with llama.cpp.
📥 Downloads (30d): 568k · 📜 License: Gemma
👤 By: Unsloth (Community) · 🎯 Task: Image-Text-to-Text
📐 Size: 12B
👤 By: Unsloth (Community) · 🎯 Task: Image-Text-to-Text
📐 Size: 12B
What it is: A GGUF-quantized version of Google's Gemma 4 12B model, optimized for local inference with llama.cpp and compatible tools. Multiple quantization levels available from 2-bit to 8-bit. Why you'd want it: Run Google's latest multimodal model on consumer hardware without needing a high-end GPU or cloud API.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on consumer hardware | Quality degrades at lower quantization |
| 568k downloads proves reliability | Still needs 8-16GB RAM minimum |
| Multiple quantization options | No audio input support in GGUF format |
unsloth/gemma-4-12b-it-GGUF · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Ideogram's latest image generation model, now available as open weights - known for best-in-class text rendering in generated images.
📥 Downloads (30d): 4.4k · 📜 License: Apache 2.0
👤 By: Ideogram AI · 🎯 Task: Text-to-Image
📐 Size: Not specified
👤 By: Ideogram AI · 🎯 Task: Text-to-Image
📐 Size: Not specified
What it is: An fp8-quantized (8-bit floating point) version of Ideogram 4, a 9.3B-parameter diffusion transformer for image generation. The nf4 variant fits on a single 24GB GPU. Why you'd want it: If you need AI-generated images with accurate text in them - signs, logos, labels, UI mockups - Ideogram consistently leads on text rendering quality.
| ✓ Pros | ✗ Cons |
|---|---|
| Best text rendering in AI images | Requires 24GB+ GPU for full model |
| Apache 2.0 for commercial use | Smaller community than Stable Diffusion |
| Open weights for local deployment | Image diversity can be limited |
ideogram-ai/ideogram-4-fp8 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
JetBrains' code-focused model that shows its reasoning - designed to explain why, not just generate what.
📥 Downloads (30d): 16.9k · 📜 License: Apache 2.0
👤 By: JetBrains · 🎯 Task: Text Generation
📐 Size: 12B
👤 By: JetBrains · 🎯 Task: Text Generation
📐 Size: 12B
What it is: A 12-billion-parameter mixture-of-experts model with only 2.5B parameters active per inference. Built specifically for code understanding, completion, and explanation with visible chain-of-thought reasoning. Why you'd want it: A code model from the company that makes the most popular Java, Python, and Kotlin IDEs - designed to integrate with their editor ecosystem and show its reasoning step by step.
| ✓ Pros | ✗ Cons |
|---|---|
| Shows reasoning, not just outputs | Narrower than general-purpose models |
| Only 2.5B active params (fast) | JetBrains ecosystem focus |
| Apache 2.0 license | Limited non-code capabilities |
JetBrains/Mellum2-12B-A2.5B-Thinking · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Product Hunt
AI Launches Today
Product Hunt's daily leaderboard for June 7, 2026 was not available at time of publication. Check Product Hunt AI for today's launches.
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | 1M |
| Anthropic | Claude Sonnet 4.6 | $3.00 | $15.00 | 1M |
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 | 200K |
| OpenAI | GPT-5.5 | $5.00 | $30.00 | 272K+ |
| OpenAI | o3 | $2.00 | N/A | N/A |
| Gemini 3.5 Flash | $1.50 | $9.00 | 1M | |
| Gemini 3.1 Pro | $2.00 | $12.00 | 1M | |
| Groq | Llama 3.3 70B | $0.59 | $0.79 | 128K |
| Groq | Llama 3.1 8B | $0.05 | $0.08 | 128K |
What this means: Anthropic and OpenAI are now price-matched at $5/M input for their flagship models (Opus 4.8 vs GPT-5.5), but Anthropic offers a 1M context window at flat rates while OpenAI charges 2x input and 1.5x output for prompts exceeding 272K tokens. Google's Gemini 3.5 Flash at $1.50/$9.00 remains the best value for high-volume work that does not require the largest models. Groq continues to offer the cheapest inference for open-source models on their custom LPU chips, with Llama 3.1 8B at 25-100x cheaper than frontier models. All providers offer batch processing at 50% off and prompt caching at up to 90% off cached input costs.
arXiv Paper of the Day
Beyond Scaling: Agents Are Heading to the Edge
Multiple authors - arXiv:2605.18535
What it claims: On-device LLM agent deployment is technically plausible today through small models, quantization, memory-aware inference, and mobile deployment frameworks - not just a future aspiration.
Key finding: Hierarchical, graph-based, and swarm-style agent organization assumptions all change fundamentally when agents run on edge devices with limited compute, requiring new coordination patterns that do not exist in cloud-scale systems.
Why practitioners should care: If edge deployment works, it eliminates per-token API costs, removes latency from network round trips, and keeps user data entirely on-device. The paper maps which current techniques (quantization, speculative decoding, memory-aware scheduling) already work on mobile hardware and which need more research.
Key finding: Hierarchical, graph-based, and swarm-style agent organization assumptions all change fundamentally when agents run on edge devices with limited compute, requiring new coordination patterns that do not exist in cloud-scale systems.
Why practitioners should care: If edge deployment works, it eliminates per-token API costs, removes latency from network round trips, and keeps user data entirely on-device. The paper maps which current techniques (quantization, speculative decoding, memory-aware scheduling) already work on mobile hardware and which need more research.
×
Click anywhere or press ESC to close
