Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
1.5
T+ parameters makes this one of the
Cursor Announces a 1.5-Trillion-Parameter Model Trained on 1
Top Story
5
leads with 1,587 Elo but satisfies rubrics
New Benchmark Reveals AI Agents Satisfy Only 3% of Criteria
5.2
ranks as the strongest open model at
New Benchmark Reveals AI Agents Satisfy Only 3% of Criteria
One Thing to Tell Your Friends
One Thing to Tell Your Friends
The company that makes the Cursor code editor just announced it's training one of the largest AI models ever built - more than 1.5 trillion parameters on 100,000 GPUs - going head-to-head with the labs that invented AI.
Summary
TL;DR
Top Stories
Trends
Tool Companies Are Becoming Model Companies, The Fable Crisis Is Reshaping AI Governance in Real Time, and Agent Benchmarking Gets Honest.
Creative AI
OpenMontage: Open and Kling 3.0: Faster Generation, Lower Costs, Full 4K.
Dev Tools
Headroom: 60 and HuggingFace Agent.
Research
AI Agents Perform Better When They Communicate in Raw Math Instead of English, Laguna M.1: A New 225B Open Coding Model Under Apache 2.0, and PreAct: Compile Successful Agent Runs Into Replayable State Machines.
Business
Hyundai Takes Full Control of Boston Dynamics as SoftBank Exits for $325 Million, Amazon Drops Sam Altman Biopic After Announcing OpenAI Partnership, and Turbopuffer Cuts Vector Database Pricing by 75%.
Surprising
VirtueBench: Testing Whether AI Has Courage, Temperance, and Justice, Scott Alexander: 20% Chance Superintelligent AI Wants to Eliminate Humanity, and The "Ask HN" Thread: Has Anyone Actually Replaced Claude with a Local Model for Daily Coding?.
Worth Watching
The Tool-to, Congress Is Writing AI Export Control Guardrails in Real Time, and AA.
GitHub
HuggingFace
Product Hunt
Top launches: Claude Code Artifacts (361), Zernio WhatsApp API (275), and Firecrawl Research Index (180).
API Pricing
What this means:** The best value for high-capability work remains GPT-4.1 at $2/$8 (input/output per million tokens) and Groq's Kimi K2 at $1/$3 for tasks that don't need frontier reasoning.
arXiv
Quantized Reasoning Models Think They Need to Think Longer, but They Do Not — A training-free logit penalty on overthinking tokens ("wait," "but," "alternatively") reduces reasoning length by 12-23% and cuts overthinking errors by up to 58% across five models from 1.5B to 32B parameters.
FYI
Hot off the Presses
01
Cursor Announces a 1.5-Trillion-Parameter Model Trained on 100,000 GPUs
What this means for you: The tools you use to write code may soon be powered by models as large as anything from OpenAI or Google - built specifically for programming.
Cursor, the AI-powered code editor with millions of users, revealed it is training a model with more than 1.5 trillion parameters across 100,000 GPUs (Graphics Processing Units - the specialized chips that power AI training). This puts a tool company in the same league as frontier AI labs in terms of raw model scale.
The move follows a pattern: Cursor's recent "/automate" feature already lets users configure agents with natural-language instructions and Slack/GitHub triggers. A proprietary model would let them optimize every interaction for coding without depending on external providers.
- 1.5T+ parameters makes this one of the largest models currently in training, rivaling GPT-5.5 and Gemini 3
- Purpose-built for coding - unlike general-purpose models, this is designed specifically for software development tasks
- Signals a shift where the companies building AI tools are no longer content to rent capabilities from labs - they want to own the intelligence layer
02
DeepSeek Raises $7.5 Billion at a $50 Billion Valuation
What this means for you: The company that proved you could build world-class AI for a fraction of the usual cost now has more money than most of its competitors.
DeepSeek, the Chinese AI lab that shocked the industry in early 2025 by building competitive models at dramatically lower cost, has raised $7.5 billion in new funding. The $50 billion valuation places it among the most valuable AI companies in the world.
""DeepSeek V4 Pro costs $0.04 per task - roughly 25x cheaper than equivalent frontier models.""
- $7.5 billion is one of the largest private AI funding rounds ever
- DeepSeek V4 Pro recently made its 75% price cut permanent, with the model now costing $0.04 per typical task
- The funding validates the "efficiency-first" approach to AI - proving that throwing more money at bigger models isn't the only path forward
03
The Fable Crisis Enters Its Second Week: Congress Responds, Experts Say the Government's Demand Is Mathematically Impossible
What this means for you: The government's demand that AI models be unable to find security flaws in code may be impossible to fulfill without breaking the models for everyone.
> Previously: June 13 - The White House imposed export controls on Anthropic's Fable 5 and Mythos 5 models, pulling them from every customer worldwide.
Today: Three significant developments mark day ten of the crisis. First, both the Senate and House have drafted FY27 NDAA (National Defense Authorization Act) provisions that would add procedural guardrails to Defense Department supply-chain authorities and explicitly bar their use as negotiation leverage - a direct rebuke of how the Fable situation was handled.
- The "jailbreak" was asking the model to fix code - the vulnerability that triggered the export controls was simply requesting Fable to identify security weaknesses, which every other frontier model (including GPT-5.5 and Opus 4.8) also does
- Zvi Mowshowitz argues the fix is mathematically impossible - you cannot distinguish between "find bugs in my code" (defensive) and "find exploits in this code" (offensive) at the classifier level without destroying the model's coding ability
- Joshua Achiam (OpenAI) warns the Fable dispute could normalize digital citizenship verification requirements across all software, creating a dangerous precedent for state control of AI access
- Odds of resolution by July 1 are "slightly under even money" according to Mowshowitz
04
New Benchmark Reveals AI Agents Satisfy Only 3% of Criteria on Real Multi-Week Projects
What this means for you: The AI agents that look impressive in demos fall apart when given the kind of work that actually matters - sustained, multi-week projects with fragmented inputs.
The AA-Briefcase benchmark puts AI agents through multi-week project evaluations with more than 1,000 fragmented inputs - the kind of messy, real-world context that actual knowledge workers deal with daily. The results are humbling.
""The best AI agent in the world completes only 3% of the criteria on tasks that resemble actual work.""
- Claude Fable 5 leads with 1,587 Elo but satisfies rubrics on only 3% of tasks
- GLM-5.2 ranks as the strongest open model at 1,266 Elo
- 1,000+ fragmented inputs simulate the reality of projects where context is scattered across documents, emails, and conversations
- The gap between demo performance and real-world performance is the largest ever measured in agentic AI
05
AI Contributed $1.26 Trillion to the Global Economy in 2025
What this means for you: AI is no longer a speculative investment - it's generating more economic value annually than the entire GDP of countries like the Netherlands or Turkey.
Liminal Capital estimates that AI contributed $1.26 trillion in annual economic value as of the end of 2025. The United States alone accounted for $878 billion of that total (70%), with a 95% confidence interval of $602 billion to $1.155 trillion.
- $1.26 trillion globally - the first credible, sourced estimate of AI's total economic contribution
- $878 billion in the US alone - roughly 3.5% of US GDP
- Growth is accelerating - the figure reflects deployment across healthcare, finance, software development, and customer service
- The number may still be conservative as it doesn't fully capture productivity gains from individual AI tool usage
$1.26
trillion globally**
$878
billion in the US alone**
Trends & Themes
Trends & Themes
Creative AI & Media
Creative AI & Media
OpenMontage: Open-Source AI Video Production With 500+ Agent Skills
What this means for you: You can now describe a video concept in plain language and get a complete, edited result - for free.
Try it: GitHub
- 12 production pipelines covering research, scripting, asset generation, editing, and rendering
- 500+ agent skills orchestrated through Claude, Cursor, and Copilot
- AGPL-3.0 license - fully open source with copyleft protections
- Works with both AI-generated visuals and real stock footage
Kling 3.0: Faster Generation, Lower Costs, Full 4K
What this means for you: AI-generated video is getting faster, cheaper, and higher resolution simultaneously.
- "Omni" mode generates full 4K video
- Improved lip-sync for talking-head content
- Lower per-generation costs and faster turnaround times
- Faster generation speeds across all quality tiers
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
AI Agents Perform Better When They Communicate in Raw Math Instead of English
What this means for you: The next generation of AI systems might work together in a "language" that humans can't read - but that makes them dramatically more effective.
A research paper covered by Two Minute Papers proposes that AI agents should communicate using raw latent representations (the internal mathematical signals models use to process information) instead of converting everything to English text. When multiple agents collaborate, translating between text and internal representations at each step loses information and slows everything down.
- Skipping the text translation step preserves information that would otherwise be lost
- Faster collaboration between agents because they don't need to encode and decode at each handoff
- Trade-off: humans can't inspect what the agents are saying to each other, raising transparency concerns
Laguna M.1: A New 225B Open Coding Model Under Apache 2.0
What this means for you: Another powerful, freely downloadable AI model designed specifically for writing code.
- 225B total parameters, 23B active using a sparse Mixture-of-Experts architecture
- 70 layers with 256 experts - optimized for agentic coding workflows
- Apache 2.0 license - no restrictions on commercial use
- Designed for long-horizon tasks where the agent needs to maintain context over extended problem-solving sessions
PreAct: Compile Successful Agent Runs Into Replayable State Machines
What this means for you: When an AI agent figures out how to do something, that knowledge can be saved and replayed 8-13x faster next time.
- 8.5x to 13x faster replay of previously successful agent workflows
- Converts trial-and-error into deterministic procedures - the agent explores once, then the solution becomes a repeatable recipe
- Addresses a key efficiency problem - agents currently re-discover solutions from scratch each time
Quantized Reasoning Models Overthink - and a Training-Free Fix Exists
What this means for you: AI models running on cheaper, compressed hardware find the right answer mid-thought, then talk themselves out of it - and a simple fix cuts this problem by more than half.
- 12-23% more reasoning tokens than needed are generated by compressed (quantized) models
- 52% of failures involve the model finding the correct answer partway through, then changing its mind
- A simple logit penalty on hesitation words ("wait," "but," "alternatively") fixes the problem without any retraining
- Works across five models from 1.5 billion to 32 billion parameters
Business & Industry
Business & Industry
Surprising
Surprising & Under-the-Radar
VirtueBench: Testing Whether AI Has Courage, Temperance, and Justice
AI researchers created a benchmark testing classical Christian virtues - prudence, justice, courage, and temperance - in AI models. Claude Fable 5 scores high on prudence and justice but only 77% on courage and 88% on temperance. The philosophical question of whether AI can or should exhibit virtues is suddenly a measurable empirical question.
Scott Alexander: 20% Chance Superintelligent AI Wants to Eliminate Humanity
In his latest timeline estimates, Scott Alexander puts a 25% chance on AGI (Artificial General Intelligence) arriving by 2027 and 50% by 2034. His median estimate for the gap between human-level AI and superhuman AI is less than 4 years. Most strikingly, he estimates a 20% chance that the first superintelligent AI would want to eliminate humanity given current safety efforts - and a 50% chance there would be a warning shot before a point of no return.
The "Ask HN" Thread: Has Anyone Actually Replaced Claude with a Local Model for Daily Coding?
A Hacker News thread asking whether anyone has successfully replaced cloud-based AI coding tools with locally-run models drew significant engagement. The answers reveal that while local models are improving rapidly, most developers still find cloud models meaningfully better for complex tasks - but the gap is closing fast enough that several commenters reported switching for privacy-sensitive work.
Norway Imposes Near-Ban on AI in Elementary Schools
Norway became one of the first countries to impose a near-complete ban on AI tools in elementary education. The policy affects schools nationwide and represents the strongest government stance yet against AI in K-12 education.
Worth Watching
Signals to Track
01
The Tool-to-Lab Pipeline Could Collapse the AI Industry's Structure
Cursor training a 1.5T model signals that the biggest AI companies of 2030 might be today's tool makers, not today's labs.
If tool companies can train competitive models while also controlling the user experience, they have both the data advantage (they see how people actually code) and the distribution advantage (they already have the users). The labs become API providers - important but commoditized. Watch whether other tool companies (Replit, Vercel, GitHub) follow Cursor's lead.
02
Congress Is Writing AI Export Control Guardrails in Real Time
The Fable crisis may produce the first US legislation that explicitly limits how the government can restrict AI deployment.
The NDAA provisions being drafted would create procedural requirements before defense supply-chain authorities can be used against AI companies. If passed, this would be the first concrete legislative response to the Fable crisis and could set the template for how future AI restrictions are handled. The bipartisan support makes passage likely.
03
AA-Briefcase May Force a Reckoning in Enterprise AI Sales
When the best agent scores 3% on real work, the gap between marketing and reality becomes a business risk.
Enterprise AI vendors have been selling agent capabilities based on demo-friendly benchmarks. AA-Briefcase's multi-week, 1,000-input evaluations are closer to what enterprise buyers actually need. If this benchmark gets adopted as a standard, some companies will need to dramatically revise their claims.
04
Agent Documentation Is a Double-Edged Sword
HuggingFace discovered that adding docs helps big models but breaks small ones - a fundamental design tension for the agent ecosystem.
The finding that Qwen3-14B went from 100% to 0% accuracy when given CLI documentation means agent builders can't just "add more context." The implication: agent-facing APIs need to be tested across model sizes, and what works for GPT-5.5 may actively harm smaller models that many users rely on.
05
PreAct's Replay Approach Could Make Agent Debugging Tractable
Compiling successful agent runs into state machines turns non-deterministic AI into deterministic procedures - at 8-13x the speed.
The biggest obstacle to trusting AI agents in production is their unpredictability. PreAct's approach - let the agent explore once, then freeze the successful path into a replayable recipe - could be the bridge between "AI that sometimes works" and "AI you can depend on."
GitHub Trending
Top Repos Today
Rank yesterday: New entry 🆕
⭐ Stars today: +3,938 · 📦 Total: 38,532
📜 License: Apache-2.0 · 👤 By: Individual developer
🎯 Time to value: 5 minutes
📜 License: Apache-2.0 · 👤 By: Individual developer
🎯 Time to value: 5 minutes
What it is: A compression tool that reduces the token count of AI agent inputs by 60-95% while preserving answer accuracy. It works as a Python/TypeScript library, proxy server, or MCP integration, using multiple compression algorithms including JSON crushing, code AST analysis, and a trained ML model. Why you'd want it: If your AI agents are hitting context limits or running up API costs, this dramatically cuts token usage without sacrificing output quality.
| ✓ Pros | ✗ Cons |
|---|---|
| 60-95% token reduction with accuracy preservation | New project - limited production battle-testing |
| Multiple integration modes (library, proxy, MCP) | Compression adds latency to each request |
| Apache-2.0 license, no restrictions | May not preserve nuance in highly technical contexts |
GitHub - chopratejas/headroom: Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 60-95% fewer tokens, same answers. Library, proxy, MCP server.
Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 60-95% fewer tokens, same answers. Library, proxy, MCP server. - chopratejas/headroom
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +1,516 · 📦 Total: 24,064
📜 License: Apache-2.0 · 👤 By: Google Research
🎯 Time to value: 15 minutes
📜 License: Apache-2.0 · 👤 By: Google Research
🎯 Time to value: 15 minutes
What it is: A pre-trained foundation model for time-series forecasting from Google Research. Built on a decoder-only architecture, it generates both point predictions and probability ranges for future values without task-specific fine-tuning. Why you'd want it: If you need to forecast trends in financial data, sensor readings, or demand planning, this gives you Google-quality predictions without training your own model.
| ✓ Pros | ✗ Cons |
|---|---|
| Works out of the box on diverse time-series data | Requires understanding of time-series data formats |
| Generates confidence intervals, not just point estimates | Large model size may be overkill for simple forecasting |
| Apache-2.0 from a major research lab | Limited to univariate forecasting in current release |
GitHub - google-research/timesfm: TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.
TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting. - google-research/timesfm
GitHub
Rank yesterday: Holding steady ➡
⭐ Stars today: +1,113 · 📦 Total: 233,307
📜 License: MIT · 👤 By: Prime Radiant (company)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: Prime Radiant (company)
🎯 Time to value: 10 minutes
What it is: A structured methodology and toolkit for AI coding agents. Instead of letting agents jump straight into coding, it guides them through design refinement, implementation planning, test-driven development, and systematic review. Think of it as engineering discipline for AI agents. Why you'd want it: If your AI coding agent produces quick-and-dirty code that needs heavy human review, this framework teaches it better engineering habits.
| ✓ Pros | ✗ Cons |
|---|---|
| Works with Claude Code, Cursor, and Copilot | Adds overhead to simple tasks that don't need full process |
| MIT license, 233K+ stars, active community | Opinionated methodology may clash with team workflows |
| Measurably reduces code review cycles | Requires initial setup and learning the framework's approach |
GitHub - obra/superpowers: An agentic skills framework & software development methodology that works.
An agentic skills framework & software development methodology that works. - obra/superpowers
GitHub
Rank yesterday: Rising ↑
⭐ Stars today: +1,055 · 📦 Total: 8,167
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 10 minutes
What it is: An MCP server that indexes your codebase into a searchable knowledge graph using tree-sitter parsing. It provides 14 specialized tools for querying code dependencies, tracing function calls, detecting dead code, and mapping architectural patterns across 158 programming languages. Why you'd want it: If your AI coding assistant wastes time re-reading files or loses track of how your codebase fits together, this gives it a persistent, queryable map of your project's architecture.
| ✓ Pros | ✗ Cons |
|---|---|
| Supports 158 languages via tree-sitter | Indexing large codebases takes significant time and memory |
| Millisecond query responses after initial indexing | Requires MCP-compatible client (Claude Code, etc.) |
| MIT license, works as background service | Knowledge graph can become stale if not re-indexed |
GitHub - DeusData/codebase-memory-mcp: High-performance code intelligence MCP server. Indexes codebases into a persistent knowledge graph — average repo in milliseconds. 158 languages, sub-ms queries, 99% fewer tokens. Single static binary, zero dependencies.
High-performance code intelligence MCP server. Indexes codebases into a persistent knowledge graph — average repo in milliseconds. 158 languages, sub-ms queries, 99% fewer tokens. Single static bin…
GitHub
Rank yesterday: Falling ↓
⭐ Stars today: +478 · 📦 Total: 4,556
📜 License: Apache-2.0 · 👤 By: Zhipu AI (research lab)
🎯 Time to value: 30 minutes
📜 License: Apache-2.0 · 👤 By: Zhipu AI (research lab)
🎯 Time to value: 30 minutes
What it is: Z.ai's 744-billion parameter open language model designed for complex coding and agentic engineering. It uses IndexShare sparse attention for efficient long-context inference up to 1 million tokens. Why you'd want it: If you want the strongest open-weight model for building autonomous coding agents or tackling complex multi-step engineering tasks.
| ✓ Pros | ✗ Cons |
|---|---|
| MIT-licensed frontier-class coding model | Requires substantial GPU infrastructure |
| 1M-token context with efficient sparse attention | Brand new - limited community tooling so far |
| Competitive with Opus 4.8 on multiple benchmarks | Quantized variants not yet independently verified |
GitHub - zai-org/GLM-5: GLM-5: From Vibe Coding to Agentic Engineering
GLM-5: From Vibe Coding to Agentic Engineering. Contribute to zai-org/GLM-5 development by creating an account on GitHub.
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +236 · 📦 Total: 6,243
📜 License: AGPL-3.0 · 👤 By: Individual developer
🎯 Time to value: 20 minutes
📜 License: AGPL-3.0 · 👤 By: Individual developer
🎯 Time to value: 20 minutes
What it is: An AI-powered end-to-end video production system with 12 pipelines and 500+ agent skills. It automates research, scripting, asset generation, editing, and rendering through AI coding assistants. Why you'd want it: If you want to produce polished video content without a production team, this lets you describe a concept in plain language and get a complete result.
| ✓ Pros | ✗ Cons |
|---|---|
| Complete pipeline from concept to rendered video | AGPL license requires sharing modifications |
| Works with multiple AI agent platforms | Complex setup with many dependencies |
| Both AI-generated and stock footage support | Quality varies significantly by input clarity |
GitHub - calesthio/OpenMontage: World’s first open-source, agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. Turn your AI coding assistant into a full video production studio.
World’s first open-source, agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. Turn your AI coding assistant into a full video production studio. - calesthio/OpenMontage
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +210 · 📦 Total: 3,891
📜 License: MIT · 👤 By: Builder.io (company)
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: Builder.io (company)
🎯 Time to value: 15 minutes
What it is: A framework for building applications that serve both human users and AI agents through dual interfaces. Instead of bolting AI onto existing UIs, it provides structured data views that agents can parse efficiently alongside the visual interface humans use. Why you'd want it: If you're building a web app that needs to work well for both human users and AI agents accessing it programmatically, this handles the dual-interface problem.
| ✓ Pros | ✗ Cons |
|---|---|
| Solves the human+agent UI problem elegantly | New framework - small ecosystem |
| MIT license from established company (Builder.io) | Requires rethinking existing UI architecture |
| Reduces agent token consumption vs scraping HTML | Limited to web applications |
GitHub - BuilderIO/agent-native: A framework for building agent-native applications.
A framework for building agent-native applications. - BuilderIO/agent-native
GitHub
Rank yesterday: Falling ↓
⭐ Stars today: +196 · 📦 Total: 12,445
📜 License: Apache-2.0 · 👤 By: Lightricks (company)
🎯 Time to value: 20 minutes
📜 License: Apache-2.0 · 👤 By: Lightricks (company)
🎯 Time to value: 20 minutes
What it is: An audio-video generative model that creates video with synchronized audio from text prompts. Includes LoRA training support for customizing the model's style and output characteristics. Why you'd want it: If you need to generate video with matching audio - for social media content, product demos, or creative projects - this handles both modalities together.
| ✓ Pros | ✗ Cons |
|---|---|
| Combined audio + video generation in one model | Requires significant GPU memory for generation |
| LoRA training for style customization | Audio quality lags behind dedicated TTS models |
| Apache-2.0 license with commercial use allowed | Generation times can be lengthy for high-quality output |
GitHub - Lightricks/LTX-2: Official Python inference and LoRA trainer package for the LTX-2 audio–video generative model.
Official Python inference and LoRA trainer package for the LTX-2 audio–video generative model. - Lightricks/LTX-2
GitHub
HuggingFace Trending
Top Models Today
The 753B open-weight model that multiple independent reviewers now call competitive with Opus 4.8 and GPT-5.5.
📥 Downloads (30d): 4,307 · 📜 License: MIT
👤 By: Z.ai (Zhipu AI) · 🎯 Task: text-generation
📐 Size: 753B (MoE)
👤 By: Z.ai (Zhipu AI) · 🎯 Task: text-generation
📐 Size: 753B (MoE)
What it is: Z.ai's flagship language model with 753 billion total parameters using a Mixture-of-Experts architecture. It features IndexShare technology that reuses indexers across sparse attention layers, cutting per-token compute by 2.9x at long contexts, and supports a 1-million-token context window. Why you'd want it: If you need frontier-class reasoning and coding in an MIT-licensed open model, GLM-5.2 scores 99.2 on AIME 2026 and 62.1 on SWE-bench Pro - competitive with proprietary models at zero licensing cost.
| ✓ Pros | ✗ Cons |
|---|---|
| MIT license with no regional restrictions | 753B parameters requires substantial GPU infrastructure |
| 1M-token context with stable long-horizon performance | Brand-new release with limited community fine-tunes |
| Near-perfect math scores (AIME 2026: 99.2) | Quantized quality unverified by third parties |
zai-org/GLM-5.2 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A 428B multimodal model that processes text, images, and video with 9x faster long-context prefill.
📥 Downloads (30d): 2,100 · 📜 License: Custom (research + commercial)
👤 By: MiniMax AI · 🎯 Task: multimodal
📐 Size: 428B/23B active
👤 By: MiniMax AI · 🎯 Task: multimodal
📐 Size: 428B/23B active
What it is: MiniMax's multimodal model that handles text, images, and video inputs through a sparse MoE architecture. Its standout feature is sparse attention achieving 9x prefill speedups at 1M context, making it practical for processing long documents or video. Why you'd want it: If you need to build applications that understand multiple types of media - text, images, and video together - in a single model with efficient long-context processing.
| ✓ Pros | ✗ Cons |
|---|---|
| True multimodal (text + image + video) | Custom license may restrict some commercial uses |
| 9x prefill speedup at 1M context | 428B total params still requires significant hardware |
| Only 23B active parameters per query | Smaller community than competing multimodal models |
MiniMaxAI/MiniMax-M3 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Moonshot AI's trillion-parameter coding specialist with 384 experts and 256K context.
📥 Downloads (30d): 1,850 · 📜 License: Apache-2.0
👤 By: Moonshot AI · 🎯 Task: text-generation (code)
📐 Size: 1T/32B active
👤 By: Moonshot AI · 🎯 Task: text-generation (code)
📐 Size: 1T/32B active
What it is: A purpose-built coding model with 1 trillion total parameters, 384 experts, and 32B active parameters per query. It supports 256K context and multimodal inputs, designed specifically for agentic software engineering workflows. Why you'd want it: If you're building AI coding agents and want a model specifically trained for sustained, multi-step software engineering rather than general chat.
| ✓ Pros | ✗ Cons |
|---|---|
| Apache-2.0 license, 1T scale for coding | Requires multi-GPU setup for full-precision inference |
| 384 experts provide deep specialization | Coding-focused means weaker at general knowledge tasks |
| Multimodal support for UI/design understanding | New release with limited benchmark verification |
moonshotai/Kimi-K2.7-Code · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Google DeepMind's model that generates text 4x faster by abandoning sequential prediction.
📥 Downloads (30d): 3,200 · 📜 License: Apache-2.0
👤 By: Google DeepMind · 🎯 Task: text-generation
📐 Size: 25.2B/3.8B active
👤 By: Google DeepMind · 🎯 Task: text-generation
📐 Size: 25.2B/3.8B active
What it is: A 25.2B-parameter model that uses discrete diffusion instead of the standard autoregressive (one-word-at-a-time) approach. It generates multiple tokens simultaneously, achieving speeds over 1,100 tokens per second with only 3.8B active parameters. Why you'd want it: If you need fast text generation and can trade some quality for 4x speed improvement - ideal for applications where response time matters more than peak quality.
| ✓ Pros | ✗ Cons |
|---|---|
| 1,100+ tokens/sec with 3.8B active params | Diffusion approach has quality trade-offs vs autoregressive |
| Apache-2.0 from Google DeepMind | Newer architecture with less community tooling |
| Only 3.8B active params - runs on consumer GPUs | Not yet proven across diverse generation tasks |
DeepSeek's 1.6T flagship with 3M+ downloads and permanent 75% price reduction.
📥 Downloads (30d): 3,100,000+ · 📜 License: MIT
👤 By: DeepSeek AI · 🎯 Task: text-generation
📐 Size: 1.6T/49B active
👤 By: DeepSeek AI · 🎯 Task: text-generation
📐 Size: 1.6T/49B active
What it is: DeepSeek's flagship model with 1.6 trillion total parameters and 49B active per query. It supports a 1-million-token context window and scores 90.1% on MMLU and 76.8% on HumanEval. Why you'd want it: The most downloaded open model on HuggingFace, combining frontier-class performance with MIT licensing and the lowest cost-per-task of any model at this capability level.
| ✓ Pros | ✗ Cons |
|---|---|
| MIT license, 3M+ monthly downloads | 1.6T total requires enterprise hardware for full model |
| $0.04 per task via API - 25x cheaper than alternatives | Chinese-developed model may face future export restrictions |
| 1M context window, strong reasoning scores | Quantized variants trade quality for accessibility |
deepseek-ai/DeepSeek-V4-Pro · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
NVIDIA's 3B visual grounding model with 2.5x throughput for object localization.
📥 Downloads (30d): 890 · 📜 License: Non-commercial
👤 By: NVIDIA · 🎯 Task: visual-grounding
📐 Size: 3B
👤 By: NVIDIA · 🎯 Task: visual-grounding
📐 Size: 3B
What it is: A 3-billion parameter model that locates objects, text, and UI elements in images. It uses Parallel Box Decoding for 2.5x throughput compared to sequential approaches, making it practical for real-time applications. Why you'd want it: If you need AI that can point to specific things in images or screenshots - useful for automated testing, accessibility tools, or robotic vision.
| ✓ Pros | ✗ Cons |
|---|---|
| 2.5x faster than sequential localization methods | Non-commercial license limits business use |
| Only 3B params - runs on consumer hardware | Specialized to localization, not general vision |
| From NVIDIA with strong computer vision heritage | Limited to static images, not video streams |
nvidia/LocateAnything-3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Cohere's 30B coding MoE that scores 67.6% on SWE-Bench Verified with only 3B active parameters.
📥 Downloads (30d): 1,200 · 📜 License: Apache-2.0
👤 By: Cohere Labs · 🎯 Task: text-generation (code)
📐 Size: 30B/3B active
👤 By: Cohere Labs · 🎯 Task: text-generation (code)
📐 Size: 30B/3B active
What it is: A coding-focused Mixture-of-Experts model with 30B total parameters but only 3B active per query. Despite its small active size, it achieves 67.6% on SWE-Bench Verified, making it one of the most efficient coding models available. Why you'd want it: If you need a capable coding model that runs on modest hardware - 3B active parameters means it can work on a laptop GPU while still solving real software engineering tasks.
| ✓ Pros | ✗ Cons |
|---|---|
| 67.6% SWE-Bench with only 3B active params | Coding-specialized, weaker at general tasks |
| Apache-2.0, runs on consumer hardware | 30B total still needs ~16GB VRAM for quantized inference |
| From Cohere with enterprise support available | Fewer experts than larger MoE models limits ceiling |
Microsoft's exploration subagent that cuts main-agent token use by 60% while improving task resolution.
📥 Downloads (30d): 780 · 📜 License: MIT
👤 By: Microsoft · 🎯 Task: text-generation (code)
📐 Size: 4B
👤 By: Microsoft · 🎯 Task: text-generation (code)
📐 Size: 4B
What it is: A small, specialized model designed to work as an "exploration subagent" - it reads and summarizes codebases so the main AI agent doesn't have to. This cuts the main agent's token consumption by 60% while actually improving resolution rates on coding tasks. Why you'd want it: If your AI coding agent burns through tokens reading files it doesn't need, this handles the exploration step cheaply and feeds only the relevant context upstream.
| ✓ Pros | ✗ Cons |
|---|---|
| 60% token reduction for main agent | Only useful as part of a multi-agent setup |
| Actually improves resolution rates | 4B model may miss nuance in complex codebases |
| MIT license, runs on minimal hardware | Requires orchestration layer to coordinate with main agent |
Product Hunt
AI Launches Today
Preview and share your coding work live as it happens
🔥 Upvotes: 361 · 👤 By: Anthropic
💰 Pricing: Freemium · 🏷 Category: AI Coding
💰 Pricing: Freemium · 🏷 Category: AI Coding
Desktop app for parallel agentic coding across multiple repos with live diff review and file editing. Lets developers preview and share coding artifacts in real-time as Claude Code generates them. Verdict: Anthropic's natural extension of Claude Code that addresses the visibility gap in agentic coding workflows.
Claude Code Desktop App Redesigned: Run parallel coding agents from one desktop workspace | Product Hunt
Claude Code’s desktop app is redesigned for parallel agentic coding. Run sessions across multiple repos, review diffs, edit files, and ship without leaving the app. Built for developers running Claude Code on Pro, Max, Team, or Enterprise.
Product Hunt
One API for WhatsApp: messaging, calling, and AI agents
🔥 Upvotes: 275 · 👤 By: Zernio
💰 Pricing: Paid · 🏷 Category: AI Agents / Messaging
💰 Pricing: Paid · 🏷 Category: AI Agents / Messaging
Unified API supporting 15 platforms for WhatsApp messaging, calling, publishing, and AI agent deployment. Verdict: Solves real developer pain around WhatsApp integration fragmentation.
Zernio: Social media API for AI agents and developers | Product Hunt
Build social media & messaging features at scale. Zernio provides a single API for publishing, inbox, analytics, ads, and account management across 15 platforms, so you can ship product instead of wasting time with platform APIs.
Product Hunt
An index for agents pushing the frontier of AI/ML research
🔥 Upvotes: 180 · 👤 By: Firecrawl
💰 Pricing: Freemium · 🏷 Category: AI Research
💰 Pricing: Freemium · 🏷 Category: AI Research
An API to search, scrape, and interact with the web at scale, specifically tuned for AI and ML research discovery. Verdict: Firecrawl continues expanding its moat in agent-web interaction.
Firecrawl: The Web Data API for AI Agents and Developers | Product Hunt
The API to search, scrape, and interact with the web at scale.
Product Hunt
Turn any API into an MCP server for AI agents
🔥 Upvotes: 172 · 👤 By: API to MCP team
💰 Pricing: Freemium · 🏷 Category: Developer Tools
💰 Pricing: Freemium · 🏷 Category: Developer Tools
Converts any REST or GraphQL API into an MCP server, enabling AI agents to interact with existing services without custom integration. Verdict: Perfectly timed for the MCP adoption wave.
API to MCP: Turn any API into an MCP server for AI agents | Product Hunt
API To MCP turns REST, GraphQL, SaaS, and internal business APIs into hosted MCP servers that AI agents can use in minutes. Build visually from the dashboard, or let an AI agent create, test, and deploy tools from API docs. End users can connect live MCP servers to ChatGPT, Claude, Codex, Cursor, VS Code, Antigravity, or custom agents with OAuth, secure auth, workflows, and forkable snapshots.
Product Hunt
Gemini-powered AI agent for insights and faster ad decisions
🔥 Upvotes: 131 · 👤 By: Google
💰 Pricing: Free · 🏷 Category: AI Analytics
💰 Pricing: Free · 🏷 Category: AI Analytics
Gemini-powered AI agent built into Google Ad Manager for natural language queries about ad performance. Verdict: Google embedding Gemini into its ad stack signals where enterprise AI agents are heading.
Ask Ad Manager by Google Ads: Gemini-powered AI agent for insights & faster ad decisions | Product Hunt
AI agent, built with Gemini, helps publishers get deeper insights, understand their performance and make better decisions faster.
Product Hunt
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Fable 5 | $10.00 | $50.00 | 1M |
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | 1M |
| Anthropic | Claude Sonnet 4.6 | $3.00 | $15.00 | 1M |
| OpenAI | GPT-5.5 | $5.00 | $30.00 | 1M |
| OpenAI | GPT-4.1 | $2.00 | $8.00 | 1M |
| OpenAI | o4-mini | $1.10 | $4.40 | 200K |
| Gemini 3.5 Flash | $1.50 | $9.00 | 1M | |
| Gemini 3.1 Pro Preview | $2.00 | $12.00 | 1M | |
| Groq | Kimi K2 Instruct | $1.00 | $3.00 | 128K |
| Groq | Llama 3.1 8B | $0.05 | $0.08 | 128K |
What this means: The best value for high-capability work remains GPT-4.1 at $2/$8 (input/output per million tokens) and Groq's Kimi K2 at $1/$3 for tasks that don't need frontier reasoning. Claude Fable 5 commands the highest pricing at $10/$50 but remains unavailable due to export controls. DeepSeek V4 Pro (not listed - API pricing is $0.14/$0.28 per million tokens) offers by far the cheapest frontier-class inference. The price floor continues to drop: Groq's Llama 3.1 8B at $0.05/$0.08 is approaching negligible cost for lightweight tasks.
arXiv Paper of the Day
Quantized Reasoning Models Think They Need to Think Longer, but They Do Not
Sanae Lotfi, Polina Kirichenko, Steven Li, Zechun Liu - arXiv:2606.00206
What it claims: When reasoning models are compressed (quantized) for cheaper deployment, they develop a pathological overthinking pattern - generating 12-23% more reasoning tokens than their uncompressed versions. Worse, 52% of their failures involve finding the correct answer mid-chain-of-thought and then talking themselves out of it.
Key finding: A training-free logit penalty on overthinking tokens ("wait," "but," "alternatively") reduces reasoning length by 12-23% and cuts overthinking errors by up to 58% across five models from 1.5B to 32B parameters.
Why practitioners should care: Anyone deploying quantized reasoning models - the standard approach for cost-efficient inference - gets an immediate, zero-training fix that saves compute and improves accuracy simultaneously. The insight that compressed models find correct answers then abandon them is directly actionable in production.
Key finding: A training-free logit penalty on overthinking tokens ("wait," "but," "alternatively") reduces reasoning length by 12-23% and cuts overthinking errors by up to 58% across five models from 1.5B to 32B parameters.
Why practitioners should care: Anyone deploying quantized reasoning models - the standard approach for cost-efficient inference - gets an immediate, zero-training fix that saves compute and improves accuracy simultaneously. The insight that compressed models find correct answers then abandon them is directly actionable in production.
×
Click anywhere or press ESC to close
