Moonshot released Kimi K3.1 Pro, a 3-trillion-parameter-class model that scores ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon on the Artificial Analysis index, Beacon open-sourced a memory layer that turns one coding agent's mistakes into lessons for Claude Code, Codex, Cursor, and 20+ other tools, and Recursive released Flint, a free 40 billion parameter coding model that runs offline at 125+ tokens a second on a MacBook Pro. And that's not all, here's everything in AI this week:
Together with Hubspot
100+ ChatGPT Prompts to Revolutionize Your Day
Supercharge your productivity with HubSpot's comprehensive guide to Chat GPT. This free resource is your fast track to AI mastery:
Industry-Specific Use Cases: 15+ real-world applications across various sectors
Productivity Guide: 21 best practices to 10x your efficiency with AI
Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
Challenge Buster: Overcome common AI hurdles with expert strategies
Plus, in-depth sections on email composition, content creation, customer support, and data analysis.
Microsoft's TRELLIS.2 is an open-source, 4 billion parameter model that turns a single image into a textured 3D model. It exports a GLB file with full materials that you can open in Unity, Unreal, or Blender.
Seconds, not hours: On an NVIDIA H100 it builds a model in about 3 seconds at 512³ resolution and about 17 seconds at 1024³.
Ready-made materials: Every model comes with PBR textures (base color, roughness, metallic, and opacity), so it looks right under real lighting in a game engine.
Free to use: The model and code are MIT-licensed, and you can run it locally or in a free demo on Hugging Face.
Try it now → https://huggingface.co/spaces/microsoft/TRELLIS.2
Hallmark is an open-source design skill from developer Nutlope at Together AI that steers Claude Code, Cursor, and Codex away from generic, AI-looking pages. You install it with one command, then use it to build, audit, or redesign any page.
Built-in taste: It picks a page structure for each brief, applies one of 21 visual themes, and checks the result against 57 "slop test" rules before handing it back.
Four modes: Build is the default. Audit lists problems in existing code without editing it, redesign rebuilds a page while keeping its copy and brand, and study pulls the layout, fonts, and colors from a screenshot or URL you like.
One-command install: Run
npx skills add nutlope/hallmark. The license is MIT, and the repo has about 25.5k GitHub stars.
Try it now → https://github.com/Nutlope/hallmark
Canva released Craft on October 22, the first image model it built itself. It debuted fourth on the DesignArena leaderboard, behind three OpenAI models, which makes Canva the second-ranked AI image lab there.
Plans before it draws: One model first maps the layout, including objects, placement, sizes, and any required text. A second model then renders the image in more than 30 aspect ratios.
2K for every user: It outputs images up to 2K and is available to all Canva users in the Dream Lab tab.
Category wins: It ranks first on DesignArena for People and Portraits and for Landscapes and Nature. Developers can call it through Canva's Connect API, and faster, cheaper versions are coming.
Try it now → https://www.canva.com
Recursive released Flint on October 21, an open-weight coding model with 39.8 billion parameters, of which about 2.5 billion are active per token. It runs offline and generates more than 125 tokens a second on an M4 Pro MacBook Pro.
Laptop-sized: The 4-bit weights are about 23 GB, so it needs a Mac or NVIDIA GPU with 48 GB of memory.
Coding scores: Recursive reports 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, and 54.9% on Terminal-Bench 2.1.
Free and open: The license is Apache 2.0, it works with Ollama, LM Studio, and llama.cpp, and a free hosted version is on OpenRouter. Recursive says its own AI system automated most of Flint's development.
Try it now → https://openrouter.ai/recursive/flint-v1.0:free
Asymptote Labs open-sourced Beacon, a memory layer that records your sessions across Claude Code, Codex, Cursor, and 20+ other agent tools. It pulls out the corrections and fixes that worked, so when one agent makes a mistake, the next one avoids it.
How it learns: Beacon captures session history, scores what worked, and extracts reusable lessons such as a correction you keep repeating or a debugging path that finally fixed a bug. You approve each lesson before agents reuse it.
Works across tools: It supports 29 local agents, including Gemini CLI, GitHub Copilot CLI, OpenCode, Cline, and Kimi Code, plus Claude.ai and ChatGPT in the browser.
Runs locally: It installs on macOS, Linux, and Windows. The license is MIT.
Try it now → https://beacon.sh
Moonshot AI released Kimi K3.1 Pro on October 23, a 3-trillion-parameter-class model with a 1 million token context window. Artificial Analysis scores it 55 on its Intelligence Index, ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon, which all score 53.
Fourth in the world: On that index, only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Meta's Muse Wonder Full (56) score higher.
Fewer made-up answers: Moonshot reports that its hallucination rate on its hardest internal test fell from 23.4% with Kimi K3 to 8.3%.
Open weights coming: Moonshot says the full weights will be released by November 4. Until then it is in the Kimi apps, Kimi Code, and the API at $2 per million input tokens and $10 per million output tokens, a third less than Kimi K3.
Try it now → https://kimi.ai
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