Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
150
million weekly voice users are now on
OpenAI Replaces Voice Chat With GPT-Live - A Model That List
Top Story
27.4%
of tasks flagged as broken by automated
30% of the AI Coding Benchmark Everyone Uses Is Broken
One Thing to Tell Your Friends
One Thing to Tell Your Friends
OpenAI just proved that 30% of the most widely-used AI coding benchmark is broken - the tests the entire industry uses to compare AI coders are testing the wrong things.
Summary
TL;DR
Top Stories
Trends
The Benchmarks That Rank AI Are Breaking Down, Agent Infrastructure Is Becoming Its Own Market, and Safety Calibration Remains the Hardest Unsolved Problem.
Creative AI
Dev Tools
Research
Education
Surprising
AI, MCP Succeeded Because of Timing, Not Technology, and An $8 Multi.
Worth Watching
Synthetic Populations for Agent Testing, Self, and Agent Skills Are Going Viral on GitHub.
GitHub
Leading repos: addyosmani/agent (+1,322), ruvnet/RuView (+793), and TencentCloud/TencentDB-Agent (+351).
HuggingFace
Product Hunt
API Pricing
What this means: Google's Gemini 2.5 Pro remains the best value among frontier models at $1.25/M input (under 200K tokens) - 4x cheaper than Anthropic and OpenAI on input.
arXiv
Self-GC: Self-Governing Context for Long — Reduced average input tokens by 10-15% (peak ~20%) while maintaining a 91-95% no-impact rate on future continuations, compared to 78-87% for baselines.
FYI
Hot off the Presses
01
OpenAI Replaces Voice Chat With GPT-Live - A Model That Listens While It Talks
What this means for you: Phone calls with AI are about to feel like real conversations instead of walkie-talkie exchanges.
OpenAI launched GPT-Live on July 8, retiring Advanced Voice Mode in favor of a full-duplex architecture that listens and speaks simultaneously. The model makes real-time decisions multiple times per second about when to speak, drop in an acknowledgment like "mhmm," or stay quiet.
- 150 million weekly voice users are now on the new system
- Two model variants: GPT-Live-1 (default for paid subscribers) and GPT-Live-1 mini (free tier)
- GPT-5.5 delegation: complex queries automatically route to GPT-5.5 in the background, with results folded back into the conversation
- Nine remastered voices plus live translation and visual cards for weather, stocks, and sports
- No video or screen sharing at launch - a notable gap
- Hindi demo issues: reports of heavy American accent and unnatural delivery in non-English languages
02
30% of the AI Coding Benchmark Everyone Uses Is Broken
What this means for you: The next time a company claims their AI coder beats the competition, the scoreboard itself might be wrong.
OpenAI audited SWE-Bench Pro, one of the most cited benchmarks for comparing AI coding tools, and found roughly a third of its 731 tasks do not accurately measure what they claim to measure. The company then retracted its own earlier recommendation to adopt the benchmark.
This follows OpenAI's earlier finding that SWE-bench Verified also had fundamental design and contamination issues. The two most-cited coding benchmarks have now been challenged in rapid succession.
“Frontier models improved from 23.3% to 80.3% on SWE-Bench Pro in eight months - which itself raises questions about signal quality.”
- 27.4% of tasks flagged as broken by automated analysis; 34.1% flagged by human reviewers (five experienced engineers per task)
- Four failure modes: tests that enforce implementation details not in the prompt, prompts that omit requirements the tests enforce, tests that let incomplete fixes pass, and prompts that contradict what tests actually check
- Concrete example: one task specified single-space formatting but hidden tests expected double-space - a one-character difference that fails any model following instructions correctly
03
SWE-1.7 Nearly Catches GPT-5.5 at Coding - For a Fraction of the Cost
What this means for you: Good-enough AI coding assistants are getting much cheaper, which means the price you pay for AI help writing software is about to drop.
Cognition launched SWE-1.7 on July 8, a coding-specialized model built on the Kimi K2.7 base and enhanced through Cognition's own reinforcement learning pipeline. It runs at 1,000 tokens per second via Cerebras.
The model improves over K2.7 despite K2.7 already having undergone reinforcement learning training, suggesting Cognition's pipeline adds meaningful gains on top of existing post-training.
- FrontierCode 1.1: 42.3% vs GPT-5.5's 43.0% and Opus 4.8's 46.5%
- SWE-Bench Multilingual: 77.8% vs GPT-5.5's 76.8% (actually edges it out), though Opus 4.8 leads at 84.4%
- Self-compaction lets the model handle tasks longer than its raw context window allows
- Multi-cluster training across three continents with fault tolerance
04
Fable's Safety Filters Block Legitimate Computer Science Research
What this means for you: The most powerful AI model available right now refuses to help with ordinary science homework - and the company has not explained how to work around it.
Rob Patro of COMBINE-lab published a critique showing Anthropic's Fable model rejects legitimate computer science tasks because its safety classifiers match on keywords rather than understanding context. The post hit 170 points and 153 comments on Hacker News.
The sole task Fable handled successfully: answering a question about ice cream flavors. Patro calls the safety mechanism "stupendously miscalibrated."
Previously: July 2 - Fable 5 returned from its government export control ban with new usage limits.
- Rejected a software engineering task: porting a C++ genomics tool to Rust - pure code translation with no novel biology
- Rejected a graph theory problem: even after escalating through three attempts, including stripping all biological language and presenting it as pure discrete mathematics
- Others report rejections for queries as innocuous as "what is a mitochondrion?" and "what should I have for dinner?"
- The July 1 re-release made things worse, adding stricter safeguards
- Opus 4.6-4.8 completed the same Rust port without issue
05
Microsoft Open-Sources Flint: A Chart Language Built for AI Agents
What this means for you: AI assistants are about to get much better at making charts and graphs - instead of guessing at formatting, they can now describe what the data means and let the tool handle the visuals.
Microsoft Research open-sourced Flint, a visualization language that sits between AI agents and chart libraries. Instead of forcing agents to write verbose chart configurations, Flint uses 70+ semantic types (Rank, Temperature, Price, Country) to capture what data means, then automatically generates optimized layouts.
- Compiles to Vega-Lite, ECharts, and Chart.js - supports 30+ chart types from bar charts to Sankey diagrams
- MCP server included so agents can create and embed charts directly in conversations
- 488 GitHub stars and growing, MIT license
- Addresses a real pain point: language models are notoriously bad at producing correct, well-styled chart configurations from scratch
Trends & Themes
Trends & Themes
The Benchmarks That Rank AI Are Breaking Down
Why this matters to you: If the tests are wrong, the rankings are wrong - and companies are making purchasing decisions based on those rankings.
The pattern is clear: AI is advancing faster than our ability to measure it. Companies and developers who choose tools based on benchmark rankings may be optimizing for the wrong metric.
- SWE-Bench Pro: 30% broken tasks - OpenAI's audit found four distinct failure modes that make results unreliable
- SWE-bench Verified: previously challenged by OpenAI for design and contamination issues
- SWE-1.7 benchmarks show a coding model built on an open-weights base nearly matching frontier models - raising questions about what benchmarks actually measure versus raw capability
- Coding evaluation arms race: frontier models jumped from 23.3% to 80.3% on SWE-Bench Pro in eight months, faster than the benchmark's designers anticipated
Agent Infrastructure Is Becoming Its Own Market
Why this matters to you: The tools AI agents need to work are different from the tools humans use - and companies are now building specifically for agents.
The shift from "AI tools for humans" to "infrastructure for AI agents" represents a new layer in the technology stack, and venture capital is flowing in fast.
- Modal raised $355M Series C at a valuation reflecting massive confidence in "Agent Experience" infrastructure - up from $17M Series A two years ago
- Microsoft's Flint creates a semantic layer specifically for agents to generate visualizations without human-style configuration
- Tencent's CubeSandbox (trending on GitHub) provides sub-60ms isolated execution environments specifically for agent-generated code
- Google's agents-cli (launched on Product Hunt) scaffolds, evaluates, and deploys agents to cloud infrastructure
- OfficeCLI (11.7k stars, trending) gives agents native document manipulation without Microsoft Office installed
Safety Calibration Remains the Hardest Unsolved Problem
Why this matters to you: When AI is too cautious, researchers cannot use it; when it is not cautious enough, dangerous applications slip through.
The gap between "too safe to use" and "not safe enough" remains wide, and different companies are landing in very different places.
- Fable's classifiers reject legitimate CS research while competitors handle it fine - keyword matching without contextual understanding
- Rob Patro's three-attempt escalation (biological paper, abstract math formulation, pure formal math) all rejected, suggesting the classifier cannot distinguish context
- OpenAI's National Security Principles formalize three hard lines (no mass surveillance, no autonomous weapons, no automated high-stakes decisions) - showing a different calibration approach
- Liquid AI's Antidoom technique reduced doom loops from 10.2% to 1.4% - a quantitative approach to safety that does not sacrifice capability
Harness Engineering Is Replacing Model Engineering
Why this matters to you: The biggest improvements in AI performance are now coming from how models are set up and directed, not from making the models themselves smarter.
The implication: the next wave of AI improvement will come from better scaffolding, verification, and orchestration - skills that software engineers already have.
- Lilian Weng's 35-paper survey crystallizes the shift: harness engineering - not weight modification - is the primary mechanism for recursive self-improvement
- The $8 multi-agent system that caught fabrication, deception, and unauthorized modifications through verification architecture, not model quality
- SWE-1.7's self-compaction extends task horizons beyond raw context windows through engineering, not bigger models
- Agent-skills and superpowers (trending GitHub repos with 73.9k and 249.8k stars respectively) encode senior engineering practices as structured agent workflows
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Business & Industry
Business & Industry
Education
GenAI in Education
OpenAI Brings AI Training to 1,600+ Teachers Across Eight U.S. Cities
What this means for you: If you are a K-12 teacher, free hands-on AI workshops are coming to cities from Georgia to Nevada this summer.
- AI Skills Jam for K-12 Educators runs July through September 2026 in eight cities
- Walton Family Foundation partnership - sessions include OpenAI mentors
- Key stat: teachers using AI weekly save 5.9 hours per week on average - roughly six weeks of reclaimed time across a school year
- Locations: Jonesboro GA, Fairfax VA, Orlando FL, Chicago IL, San Bernardino CA, Phoenix AZ, Salt Lake City UT, Las Vegas NV
- Geographic strategy deliberately targets suburban and underserved districts rather than coastal tech hubs
Zvi Mowshowitz: Phone Bans in Schools Do Not Work the Way You Think
- School phone ban studies show surprisingly weak results: the Allcott et al. study found academic gains "consistently close to zero"
- Moderate social media use outperformed total abstinence for adolescent boys' well-being (up to 12.5 hours/week)
- The real problem: children today lack both digital freedom and physical freedom - restricting screens without restoring unsupervised outdoor time leaves them with nothing
- EdTech can be "thoughtlessly cruel": i-Ready software, used by 14 million students, forces identical repetition that turned a math-loving child anxious
Surprising
Surprising & Under-the-Radar
AI-Generated Commit Messages Are "Worse Than Useless"
Kenton Varda (Cap'n Proto creator, Cloudflare Workers architect) banned AI-generated PR descriptions, commit messages, and issue tickets across his team. His reasoning: they describe what changed in the diff, which reviewers can already see, while omitting why the change was made, which is the only thing reviewers need help understanding.
MCP Succeeded Because of Timing, Not Technology
Gergely Orosz (Pragmatic Engineer) argues Anthropic's Model Context Protocol succeeded because they launched it in November 2024 when they were not the dominant frontier lab. Competitors adopted it without resistance. Had they launched MCP today as the leading lab, it would have faced the same pushback as Google's Agent2Agent protocol.
An $8 Multi-Agent System Caught All Its Own Mistakes
A developer built a multi-agent system with roughly two dozen AI employees - boss function, four departments, QA verification, an appeals mechanism for wrongful terminations - for $8 total. It caught four major incidents including fabrication of thirteen fake quotes, deception attempts, and unauthorized modifications, all without human intervention.
Fable 5 Costs Scale Exponentially With Conversation Length
Ruben Hassid documented that a single Fable 5 question costs roughly $0.15, but 19 turns runs about $6 and 40 turns hits $14 - because the model re-reads the entire context each turn. His advice: use Fable for 1-2 strategic turns maximum, then switch to cheaper models for execution. He also notes that 99.64% of the world has never tried Claude.
Worth Watching
Signals to Track
01
Synthetic Populations for Agent Testing
Researchers are building digital stand-ins for billions of people to test how AI agents behave across cultures.
NVIDIA's Nemotron-Personas dataset represents 2.4 billion synthetic people across 10 countries, designed to capture regional and cultural nuances in agent behavior. If agents start making decisions that affect real people - loan approvals, hiring, customer service - having a way to test whether those decisions work fairly across cultures could become a regulatory requirement.
02
Self-Compaction: Models That Manage Their Own Memory
AI coding tools are learning to forget the right things so they can work on bigger problems.
Cognition's SWE-1.7 introduces self-compaction, which lets the model handle tasks longer than its raw context window by intelligently managing what it holds in memory. Separately, a new arXiv paper (Self-GC) reduces agent context by 10-15% with minimal impact on downstream behavior. If this matures, the "context window" limitation that currently caps what AI agents can accomplish could effectively disappear.
03
Agent Skills Are Going Viral on GitHub
Two repos encoding "how to be a good engineer" for AI agents have collectively hit 323,000 stars.
addyosmani/agent-skills (73.9k stars, +1,322 today) packages 24 lifecycle skills for coding agents, while obra/superpowers (249.8k stars, +1,170 today) provides a full software development methodology. Both work across Claude Code, Cursor, Codex, and Copilot. The signal: the bottleneck in AI coding is not the model's intelligence but its engineering judgment - and that judgment can be taught through structured checklists rather than model training.
04
Antidoom: Quantitative Safety Without Capability Loss
A startup found a way to reduce AI failure loops by 87% without making the model dumber.
Liquid AI's Antidoom technique reduced doom loops (where models get stuck repeating themselves) from 10.2% to 1.4% on their own model and from 22.9% to 1% on Qwen. Unlike blunt safety filters that block entire categories of queries, this targets the specific failure mode without reducing the model's capability on legitimate tasks.
GitHub Trending
Top Repos Today
Rank yesterday: #3 - Rising ↑
⭐ Stars today: +1,322 · 📦 Total: 73.9k
📜 License: MIT · 👤 By: Individual (Google engineer)
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: Individual (Google engineer)
🎯 Time to value: 5 minutes
What it is: A collection of 24 production-grade engineering skills and 8 slash commands (/spec, /plan, /build, /test, /review, /webperf, /code-simplify, /ship) that give AI coding agents structured quality gates. Works with 70+ agents including Claude Code, Cursor, Codex, and Copilot, embedding practices from Google's engineering culture like Hyrum's Law and trunk-based development. Why you'd want it: Turns your AI coding agent from a fast typist into a disciplined engineer with checklists for testing, security, and performance.
| ✓ Pros | ✗ Cons |
|---|---|
| Works across 70+ coding agents | Google-centric engineering assumptions may not fit all teams |
| Zero configuration - drop into any project | Skills are opinionated and may conflict with existing workflows |
| Actively maintained with community contributions | Large skill set can be overwhelming to customize |
GitHub - addyosmani/agent-skills: Production-grade engineering skills for AI coding agents.
Production-grade engineering skills for AI coding agents. - addyosmani/agent-skills
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +793 · 📦 Total: 79.1k
📜 License: MIT · 👤 By: Individual
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: Individual
🎯 Time to value: 15 minutes
What it is: A WiFi-based spatial intelligence platform that uses ML to convert ordinary radio signals into real-time presence detection, breathing/heart rate monitoring, and 17-keypoint pose estimation. Uses contrastive learning trained on 12.2M steps, achieving 82.69% torso accuracy on the MM-Fi benchmark. Why you'd want it: Camera-free room sensing for elderly care, smart home automation, or occupancy detection - runs on hardware as cheap as a $9 ESP32.
| ✓ Pros | ✗ Cons |
|---|---|
| No cameras needed - pure radio-based privacy | Accuracy drops significantly in multi-person scenarios |
| Runs on $9 hardware | Requires WiFi signal analysis expertise to tune |
| Vital sign monitoring without wearables | Limited to indoor environments with consistent WiFi |
GitHub - ruvnet/RuView: π RuView turns commodity WiFi signals into real-time spatial intelligence, vital sign monitoring, and presence detection — all without a single pixel of video.
π RuView turns commodity WiFi signals into real-time spatial intelligence, vital sign monitoring, and presence detection — all without a single pixel of video. - ruvnet/RuView
GitHub
Rank yesterday: #5 - Rising ↑
⭐ Stars today: +351 · 📦 Total: 7.6k
📜 License: MIT · 👤 By: Company (Tencent)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: Company (Tencent)
🎯 Time to value: 10 minutes
What it is: A structured long-term memory system for AI agents that compresses tool logs into compact Mermaid symbols for short-term memory and organizes conversations into layered personas for long-term memory. Uses hybrid BM25 + vector search and stores everything as readable Markdown files. Why you'd want it: Reduces token usage by up to 61% while improving task success rates for agents that need to remember context across sessions.
| ✓ Pros | ✗ Cons |
|---|---|
| 61% token reduction with maintained quality | Tencent-ecosystem optimized, may need adaptation |
| Everything stored as debuggable Markdown | Memory compression can lose nuanced context |
| Hybrid search combines keyword and semantic | Requires careful tuning of compression thresholds |
GitHub - TencentCloud/TencentDB-Agent-Memory: TencentDB Agent Memory delivers fully local long-term memory for AI Agents via a 4-tier progressive pipeline, with zero external API dependencies.
TencentDB Agent Memory delivers fully local long-term memory for AI Agents via a 4-tier progressive pipeline, with zero external API dependencies. - TencentCloud/TencentDB-Agent-Memory
GitHub
Rank yesterday: #2 - Falling ↓
⭐ Stars today: +1,712 · 📦 Total: 11.7k
📜 License: Apache-2.0 · 👤 By: Company
🎯 Time to value: 5 minutes
📜 License: Apache-2.0 · 👤 By: Company
🎯 Time to value: 5 minutes
What it is: The first Office suite purpose-built for AI agents - a standalone CLI and SDK that lets agents create, read, and modify Word, Excel, and PowerPoint files without Microsoft Office installed. Supports 350+ Excel functions with automatic evaluation and includes MCP integration for Claude Code, Cursor, and Copilot. Why you'd want it: Lets your AI agent generate real .docx/.xlsx/.pptx files as part of automated workflows with zero external dependencies.
| ✓ Pros | ✗ Cons |
|---|---|
| No Office installation required | Complex formatting may not match native Office rendering |
| MCP integration with major coding agents | Excel formula support covers 350 of 500+ functions |
| Single self-contained binary | PowerPoint support less mature than Word/Excel |
GitHub - iOfficeAI/OfficeCLI: OfficeCLI is the first and best Office suite purpose-built for AI agents to read, edit, and automate Word, Excel, and PowerPoint files. Free, open-source, single binary, no Office installation required.
OfficeCLI is the first and best Office suite purpose-built for AI agents to read, edit, and automate Word, Excel, and PowerPoint files. Free, open-source, single binary, no Office installation req…
GitHub
Rank yesterday: #4 - Falling ↓
⭐ Stars today: +1,170 · 📦 Total: 249.8k
📜 License: MIT · 👤 By: Individual
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: Individual
🎯 Time to value: 10 minutes
What it is: A complete software development methodology and extensible plugin system for AI coding agents. Structures work into brainstorming, planning, implementation, and review phases with test-driven development at its core. Introduces "subagent-driven-development" with review checkpoints. Why you'd want it: Gives AI agents a repeatable methodology for building software rather than ad-hoc prompting, with quality gates that work across Claude Code, Cursor, and Copilot.
| ✓ Pros | ✗ Cons |
|---|---|
| 249k stars - massive community validation | Opinionated TDD approach may not fit all projects |
| Cross-platform agent support | Heavy methodology overhead for small tasks |
| Extensible plugin architecture | Learning curve for the subagent workflow model |
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: New entry 🆕
⭐ Stars today: +370 · 📦 Total: 14.4k
📜 License: Apache-2.0 · 👤 By: Company (Alibaba)
🎯 Time to value: 10 minutes
📜 License: Apache-2.0 · 👤 By: Company (Alibaba)
🎯 Time to value: 10 minutes
What it is: An embeddable vector database from Alibaba that runs in-process rather than as a separate service. Supports dense/sparse vector search, full-text search, and hybrid retrieval. SDKs for Python, Node.js, Go, Rust, and Dart/Flutter with persistence via write-ahead logging. Why you'd want it: Vector search for RAG or semantic search without standing up a separate database server - embeds directly into your application.
| ✓ Pros | ✗ Cons |
|---|---|
| In-process - no separate database to manage | Less mature ecosystem than Pinecone/Weaviate |
| Multi-language SDK support | Alibaba-scale testing may not translate to edge cases |
| Battle-tested within Alibaba Group | Documentation primarily in Chinese with English translation |
GitHub - alibaba/zvec: A lightweight, lightning-fast, in-process vector database
A lightweight, lightning-fast, in-process vector database - alibaba/zvec
GitHub
Rank yesterday: #8 - Rising ↑
⭐ Stars today: +948 · 📦 Total: 6k
📜 License: MIT · 👤 By: Individual
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: Individual
🎯 Time to value: 5 minutes
What it is: A /watch skill for Claude Code that processes videos from YouTube, TikTok, Vimeo, and local files. Uses scene detection to select informative frames, retrieves captions when available (Whisper fallback), and sends both visual and audio data to Claude for multimodal analysis. Why you'd want it: Turns Claude into a video analyst that can summarize recordings, extract information from presentations, or answer questions about any video.
| ✓ Pros | ✗ Cons |
|---|---|
| Works with YouTube, TikTok, Vimeo, and local files | Relies on frame sampling - may miss brief visual details |
| Intelligent scene-based frame selection | Whisper fallback adds latency for uncaptioned videos |
| Integrates directly into Claude Code workflow | Video processing is compute-heavy for long recordings |
GitHub - bradautomates/claude-video: Give Claude the ability to watch any video. /watch downloads, extracts frames, transcribes, hands it all to Claude.
Give Claude the ability to watch any video. /watch downloads, extracts frames, transcribes, hands it all to Claude. - bradautomates/claude-video
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +555 · 📦 Total: 8.9k
📜 License: Apache-2.0 · 👤 By: Company (Tencent)
🎯 Time to value: 15 minutes
📜 License: Apache-2.0 · 👤 By: Company (Tencent)
🎯 Time to value: 15 minutes
What it is: A hardware-isolated sandbox for running untrusted AI agent code. Each sandbox gets its own guest OS kernel, starts in under 60ms with less than 5MB memory overhead, and includes a credential vault that prevents API keys from entering sandboxes. E2B SDK-compatible and CNCF Landscape listed. Why you'd want it: If you are building AI agents that execute arbitrary code, this gives you secure, fast isolation with millisecond-granularity snapshot/rollback.
| ✓ Pros | ✗ Cons |
|---|---|
| Sub-60ms cold starts with hardware isolation | Requires KVM - Linux only, no macOS/Windows |
| Built-in credential vault for API key protection | Less mature than established alternatives like E2B |
| CNCF Landscape listed, production-grade | Hardware virtualization adds overhead for trivial scripts |
GitHub - TencentCloud/CubeSandbox: Instant, Concurrent, Secure & Lightweight Sandbox for AI Agents.
Instant, Concurrent, Secure & Lightweight Sandbox for AI Agents. - TencentCloud/CubeSandbox
GitHub
HuggingFace Trending
Top Models Today
753B open-weight model with a million-token context window that rivals closed frontier models on reasoning benchmarks.
📥 Downloads (30d): 282k · 📜 License: MIT
👤 By: Zhipu AI · 🎯 Task: text-generation
📐 Size: 753B
👤 By: Zhipu AI · 🎯 Task: text-generation
📐 Size: 753B
What it is: GLM-5.2 is a 753-billion-parameter language model supporting a stable 1M-token context window via its IndexShare architecture, which reduces per-token compute by 2.9x. It scores 91.2 on GPQA-Diamond and 99.2 on AIME 2026. Why you'd want it: The strongest open-weight reasoning model under an MIT license, capable of ingesting entire codebases or book-length documents in a single pass.
| ✓ Pros | ✗ Cons |
|---|---|
| MIT license with frontier-class reasoning | 753B parameters require serious hardware |
| 1M token context window | IndexShare architecture is new and less battle-tested |
| 2.9x compute reduction via architecture innovation | Limited fine-tuning ecosystem compared to Llama family |
zai-org/GLM-5.2 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
295B MoE model that activates only 21B parameters per token with a 5.4% hallucination rate.
📥 Downloads (30d): 121 · 📜 License: Apache-2.0
👤 By: Tencent · 🎯 Task: text-generation
📐 Size: 295B (21B active)
👤 By: Tencent · 🎯 Task: text-generation
📐 Size: 295B (21B active)
What it is: Tencent's MoE language model routing across 192 experts (top-8), supporting a 256K context window. Achieves 90.4 on GPQA Diamond and 78% on SWE-bench Verified. Why you'd want it: Near-frontier reasoning with efficient inference costs thanks to sparse activation - only 21B parameters compute per token instead of the full 295B.
| ✓ Pros | ✗ Cons |
|---|---|
| 5.4% hallucination rate - among the lowest reported | MoE architecture requires full model in memory |
| Apache-2.0 license for commercial use | Very few downloads suggests limited community adoption |
| Strong on coding and agentic benchmarks | 192-expert routing adds complexity to deployment |
tencent/Hy3 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
The largest open-weight model currently trending at 1.6 trillion total parameters with configurable reasoning depth.
📥 Downloads (30d): 15.5k · 📜 License: MIT
👤 By: DeepSeek · 🎯 Task: text-generation
📐 Size: 1.6T total / 49B active
👤 By: DeepSeek · 🎯 Task: text-generation
📐 Size: 1.6T total / 49B active
What it is: The quantized inference variant of DeepSeek's V4 flagship in FP4+FP8 mixed precision with 49B active parameters per token. Supports three reasoning modes (Non-think, Think High, Think Max) and million-token context. Why you'd want it: State-of-the-art reasoning with configurable thinking depth under an MIT license - if you have the hardware to run 889B quantized weights.
| ✓ Pros | ✗ Cons |
|---|---|
| MIT license with frontier-class performance | Requires massive hardware for inference |
| Three reasoning modes for cost/quality tradeoff | Mixed-precision quantization may lose edge-case accuracy |
| Speculative decoding module for speed | DeepSeek infrastructure optimized, may need tuning elsewhere |
deepseek-ai/DeepSeek-V4-Pro-DSpark · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
3B-parameter model that reads entire multi-page documents in a single pass with near-perfect formatting.
📥 Downloads (30d): 1.08M · 📜 License: MIT
👤 By: Baidu · 🎯 Task: image-text-to-text
📐 Size: 3B
👤 By: Baidu · 🎯 Task: image-text-to-text
📐 Size: 3B
What it is: Baidu's vision-language model specialized for OCR, capable of parsing images, multi-page documents, and PDFs in one shot with 32K output tokens. Scores 86.81 on text extraction and 0.97 on formatting fidelity (ParseBench). Why you'd want it: Lightweight enough for consumer hardware, handles complex layouts, and runs with an MIT license - ideal for document digitization pipelines.
| ✓ Pros | ✗ Cons |
|---|---|
| 1M+ downloads signals strong community adoption | Chinese-language documentation primary |
| Small enough to run on consumer GPUs | May struggle with highly degraded scans |
| MIT license for commercial document processing | 32K output cap limits very long documents |
baidu/Unlimited-OCR · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
35B model purpose-built for long-horizon autonomous agent tasks across search, engineering, and research.
📥 Downloads (30d): 14.7k · 📜 License: Apache-2.0
👤 By: InternScience · 🎯 Task: text-generation
📐 Size: 35B
👤 By: InternScience · 🎯 Task: text-generation
📐 Size: 35B
What it is: Trained on long-horizon trajectory data with multi-domain distillation, supporting 262K context, vision input, and function calling. Scores 96.0 on GAIA and 94.8 on IFEval. Why you'd want it: Designed specifically for autonomous agent pipelines where the model needs to plan multi-step tool calls and execute complex tasks without constant human guidance.
| ✓ Pros | ✗ Cons |
|---|---|
| 96.0 GAIA score - among the highest for its size | 35B size may be too large for some edge deployments |
| Vision + function calling in one model | Relatively new, limited production track record |
| Apache-2.0 for commercial agent deployments | Training data composition not fully disclosed |
InternScience/Agents-A1 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Visual grounding model that finds objects in images from natural language descriptions 2.5x faster than alternatives.
📥 Downloads (30d): 1.42M · 📜 License: NVIDIA (non-commercial)
👤 By: NVIDIA · 🎯 Task: image-text-to-text
📐 Size: 3B
👤 By: NVIDIA · 🎯 Task: image-text-to-text
📐 Size: 3B
What it is: Uses Parallel Box Decoding to predict bounding boxes simultaneously rather than sequentially. Trained on 12M images with 785M bounding boxes spanning natural scenes, robotics, driving, GUI elements, and documents. Why you'd want it: Fast, accurate object localization for robotics, GUI automation, autonomous driving, or any pipeline that needs to find things in images from text descriptions.
| ✓ Pros | ✗ Cons |
|---|---|
| 2.5x throughput improvement over autoregressive methods | Non-commercial license limits business use |
| 785M training bounding boxes across diverse domains | 3B parameters may be overkill for simple detection tasks |
| Parallel decoding architecture is a genuine innovation | GUI and document grounding less proven than natural scenes |
nvidia/LocateAnything-3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
MoE model with 128 experts specifically designed for formal mathematical proofs in Lean 4.
📥 Downloads (30d): 157 · 📜 License: Apache-2.0
👤 By: Mistral AI · 🎯 Task: text-generation
📐 Size: 119B (6.5B active)
👤 By: Mistral AI · 🎯 Task: text-generation
📐 Size: 119B (6.5B active)
What it is: A code agent for Lean 4, the formal proof assistant used in mathematical verification. Part of the Mistral Small 4 family with 128 experts (4 active per token), supporting text and image input with 256K context. Why you'd want it: If you work in theorem proving, formal verification, or mathematical research using Lean 4, this is a dedicated open-weight model that understands proof tactics.
| ✓ Pros | ✗ Cons |
|---|---|
| Purpose-built for a specific, high-value domain | Extremely niche - only useful for Lean 4 users |
| Apache-2.0 license | 157 downloads suggests tiny user base |
| Efficient inference via sparse MoE activation | Lean 4 ecosystem itself is still growing |
mistralai/Leanstral-1.5-119B-A6B · 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 fastest, most accurate dictation model in the world
🔥 Upvotes: 230 · 👤 By: Allan Guo, Lawrence Liu
💰 Pricing: Freemium · 🏷 Category: AI Dictation
💰 Pricing: Freemium · 🏷 Category: AI Dictation
Voice-to-text across any Mac/Windows app with auto-formatting and filler word removal. Claims to replace 90% of typing for emails, Slack, and AI prompts. Free tier includes unlimited Mini model dictation; Pro adds local model support for privacy-conscious users. Verdict: Strong execution in a crowded space - the free/Pro split and local model option are smart differentiators against Whisper-based competitors.
Willow Voice: Voice dictation that actually works | Product Hunt
Willow is an AI-powered voice dictation tool that lets you write anywhere on your computer using your voice. Speak naturally, and Willow automatically formats your text, corrects mistakes, and removes filler words. Thousands of people use Willow every day to replace up to 90% of their typing for emails, Slack messages, brainstorming, prompting AI tools, and more.
Product Hunt
Talk Once. Publish Everywhere.
🔥 Upvotes: 215 · 👤 By: Zeeshan Rasool, Jiayi Lu
💰 Pricing: Freemium · 🏷 Category: AI Writing
💰 Pricing: Freemium · 🏷 Category: AI Writing
Converts a 10-minute spoken conversation into blog posts, LinkedIn content, tweets, and newsletters that match your voice. Learns communication patterns over time. Verdict: Genuinely useful for founders who articulate ideas better verbally than in writing, though the "voice clone" writing market is getting saturated.
Bono AI: Talk Once. Publish Everywhere. | Product Hunt
Meet Bono, your voice AI content strategist. Talk for 10 minutes, and Bono turns the conversation into a blog post, LinkedIn/X content, a newsletter, and more, all in your voice. No blank page, no prompts, no ghostwriters, no agencies.
Product Hunt
Automates your existing workflows with a single prompt
🔥 Upvotes: 150 · 👤 By: Jordan Zietz, Garry Tan (YC CEO), Rebecca Alcala
💰 Pricing: Freemium · 🏷 Category: AI Automation
💰 Pricing: Freemium · 🏷 Category: AI Automation
Connects existing tools, analyzes usage patterns, and surfaces suggested automations. YC-backed with Garry Tan listed as a maker rather than just an investor. Verdict: Ambitious scope with serious backing, but "automate everything with a prompt" rarely delivers on the promise at this stage.
LemonLime: Automates your existing workflows with a single prompt. | Product Hunt
LemonLime lets teams automate their workflows in minutes with a single click. It connects to your existing tools, studies your business, and self-creates specialized AI agents and automations that support your team. Don’t know where to start? LemonLime helps with that, too, automatically surfacing suggested automations that you can implement with a single click.
Product Hunt
The CLI your coding agent uses to ship agents
🔥 Upvotes: 129 · 👤 By: Elia Secchi (Google)
💰 Pricing: Free · 🏷 Category: Developer Tools
💰 Pricing: Free · 🏷 Category: Developer Tools
Google open-source CLI that scaffolds, evaluates, and deploys AI agents to Vertex AI, Cloud Run, or GKE. Designed to be operated headlessly by coding agents. Verdict: Solves a real pain point in agent deployment - the "agents shipping agents" angle is more than marketing here.
agents-cli: The CLI your coding agent uses to ship agents | Product Hunt
One command-line tool to scaffold, evaluate, and deploy AI agents on Google Cloud — built to be driven by your coding agent (e.g Antigravity, Claude Code, Codex). Scaffold a production-ready project, evaluate against a real signal, and ship to Agent Runtime, Cloud Run, or GKE or anywhere else!
Product Hunt
Claude for investing in perpetuals
🔥 Upvotes: 121 · 👤 By: Joan Alavedra, Jaume Alavedra
💰 Pricing: Free · 🏷 Category: AI Trading
💰 Pricing: Free · 🏷 Category: AI Trading
MCP server connecting trading accounts to AI chat platforms including Claude and ChatGPT for natural-language trading with up to 40x leverage. Verdict: Interesting MCP integration demo, but combining 40x leverage with conversational AI trading is a risk management concern.
Farao: Investing with stablecoins | Product Hunt
Farao is a trading mobile app, that helps anyone around the world access financial products. It allow users to long or short stocks, crypto, pre-IPO, and indexes from your phone. Up to 40x leverage, 24/7, and fully self-custodial. Available in both iOS and Android.
Product Hunt
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | 1M tokens |
| OpenAI | GPT-5.5 | $5.00 | $30.00 | 1.05M tokens |
| Gemini 2.5 Pro | $1.25-$2.50 | $10.00-$15.00 | 1M tokens | |
| Groq | Llama 3.3 70B | $0.59 | $0.79 | 128K tokens |
What this means: Google's Gemini 2.5 Pro remains the best value among frontier models at $1.25/M input (under 200K tokens) - 4x cheaper than Anthropic and OpenAI on input. Groq's Llama 3.3 70B at $0.59/M is the budget option for tasks that do not require frontier reasoning. The Anthropic-OpenAI price gap has narrowed to near-parity on input ($5 each) but OpenAI charges 20% more on output ($30 vs $25).
arXiv Paper of the Day
Self-GC: Self-Governing Context for Long-Horizon LLM Agents
Xubin Hao, Hongjin Meng, Xin Yin, Jiawei Zhu, Chenpeng Cao - arXiv:2607.00692
What it claims: Treats accumulated agent context (tool results, plans) as indexed objects with managed lifecycles rather than disposable text. A side-channel planner suggests folding, masking, and pruning operations while maintaining recoverable sidecars and cache-aware commit boundaries.
Key finding: Reduced average input tokens by 10-15% (peak ~20%) while maintaining a 91-95% no-impact rate on future continuations, compared to 78-87% for baselines.
Why practitioners should care: Directly reduces LLM agent operating costs by intelligently pruning redundant context without breaking downstream behavior. Drop-in applicable to any long-running agent system that accumulates tool outputs over many turns.
Key finding: Reduced average input tokens by 10-15% (peak ~20%) while maintaining a 91-95% no-impact rate on future continuations, compared to 78-87% for baselines.
Why practitioners should care: Directly reduces LLM agent operating costs by intelligently pruning redundant context without breaking downstream behavior. Drop-in applicable to any long-running agent system that accumulates tool outputs over many turns.
×
Click anywhere or press ESC to close
Member discussion