
Thinking Machines released Inkling 2 Mini, a free 20 GB model that runs offline on a laptop and scores 79.2% on SWE-bench Verified, rea-video breaks down any video shot by shot and tells you how to remake it, and tiny-voca fits a full text-to-speech model into 4.5 million parameters. And that's not all, here's everything in AI this week:
Together with Hubspot
Turn AI into Your Income Engine
Ready to transform artificial intelligence from a buzzword into your personal revenue generator?
HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.
Inside you'll discover:
A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential
Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background
Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve
Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.
Google Research released the code for PaperOrchestra, a multi-agent system that turns rough ideas, experiment logs, and a conference template into a full LaTeX research paper, with a literature review, citations, and figures.
A team of agents: Separate agents write the outline, search the literature, make plots and diagrams, draft each section, and then refine the whole paper.
Real citations: The literature agent finds papers through web search and the Semantic Scholar API and verifies them, so the reference list is built from papers that exist.
Free to use: It is Apache 2.0, the paper behind it was accepted at COLM 2026, and the repo gained more than 1,500 GitHub stars in a single day this week.
Try it now → https://github.com/google-research/paper-orchestra
Inworld released Voice Isolator on November 18, its first open-source model. It removes background chatter, music, and noise from speech while keeping the voice as it was recorded, including its tone, breaths, and laughter.
Two versions: One model cleans up recorded files, and a second one works on live audio as it streams. Both run at 48 kHz.
Beats other open models: On Inworld's own VIBE benchmark it scores 0.768, ahead of the best other open-source denoiser it tested at 0.695.
Free for commercial use: Code and weights are Apache 2.0 on GitHub and Hugging Face. Inworld also open-sourced SAVA, a model that detects when someone has finished speaking.
Try it now → https://github.com/inworld-ai/voice-isolator
OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on November 17. Astra is its frontier model for the hardest projects, and Sol is the cheaper everyday model. OpenAI says both finish demanding tasks with fewer output tokens, so you get more done before hitting your usage limits.
Astra: OpenAI says it sets new highs on professional and scientific work at less than a third of the cost per task of GPT-6 Astra. In Codex, early testers saw cleaner code and more focused tests.
Sol: On Cognition's FrontierCode 2.0, it scored 43.8% to Astra's 44.2%, at about a fifth of the cost per run.
Where to get it: Both are live in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and in the API. Astra's API price drops 20% to $8 per million input tokens and $40 per million output tokens, and Sol stays at $2 and $10.
Try it now → https://chatgpt.com/codex
Developer Muhammed Nazeem released tiny-voca, a small English text-to-speech model trained from scratch. The whole model, including the part that produces the sound wave, has 4.5 million parameters and takes up 18 MB.
Fast on a CPU: It runs about 70 times faster than real time on four CPU threads, and it can also run in your browser.
Two voices: It has one female and one male American English voice, at 24 kHz.
Cheap and open: He trained it for $9.56 on one rented RTX 4090, using about 18 hours of synthetic speech. The code, weights, training recipe, and dataset are all released under Apache 2.0.
Try it now → https://tiny-voca.n4ze3m.com
rea-video is an open-source tool for coding agents like Claude Code and Codex. Give it a video and it splits it into shots, tells you whether each one was filmed, AI-generated, or built in code, and writes a plan to remake it.
A full breakdown: It also explains how the captions, cuts, and audio were made, for talking heads, product videos, AI shorts, 3D explainers, and motion graphics.
A remake plan: You get a plan split into what to film, what to generate, and what to code, with a reverse-engineered prompt for each AI shot.
Free to use: It is Apache 2.0. You install it by asking your agent to install the repo and break down your video. The accuracy figures it reports are the author's own.
Try it now → https://github.com/JuneYaooo/rea-video
Mira Murati's Thinking Machines Lab released Inkling 2 Mini on November 20, with open weights on Hugging Face. It takes about 20 GB of memory, small enough to run offline on a laptop, and it is built for coding agents, tool use, and knowledge work.
Strong coding scores: The lab reports 79.2% on SWE-bench Verified (bash only), ahead of Qwen3.8-27B at 77.8%, and 48.2% on DeepSWE.
Small model, extra passes: It has 26.4 billion parameters, with 3.4 billion active. Its middle 12 layers run three times for every token, which adds computation without adding weights. It accepts text and images.
What you need: Ollama supports it on Macs with at least 32 GB of RAM, and it also runs through llama.cpp, vLLM, and SGLang. Hugging Face lists the license as Apache 2.0, and you can also try it on Thinking Machines' Tinker.
Thanks for making it to the end! I put my heart into every email I send. I hope you are enjoying it. Let me know your thoughts so I can make the next one even better.
See you tomorrow :)
Dr. Alvaro Cintas
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial
✓ $500+ worth of AI tools







