DeepSeek open-sourced V4.1 Flash, a 552B model it puts ahead of many top closed models for free, Meta launched Muse, a personal agent that runs your email, calendar, and payments for you, and memanto gave coding agents permanent memory with one pip install. And that's not all, here's everything in AI this week:
Meta just released Muse, a personal AI agent that connects to your email, calendar, and payments and carries out the whole task instead of just drafting it. Each user gets a dedicated, isolated cloud machine where the agent and a live browser run, and you can watch it work.
Its own secure cloud machine: Muse gets a dedicated virtual machine where the agent, its browser, and your credentials stay isolated, and Meta says it never sees your actual passwords or payment details.
Takes real actions: It can open a browser, fill out forms, book travel, and negotiate bills for you, then keeps working after you close the app and returns only when it needs your approval.
Powered by Muse Spark: The agent runs on Meta's Muse Spark model, live now in the US on iOS, Android, muse.ai, and WhatsApp, with a free tier plus $20 and $100 plans.
Try it now → https://muse.ai
Gawkbot is a new open-source tool that turns your existing AI model into an always-on workflow bot. Describe a job in one sentence, or demo it once on a screen share, and it builds the bot that runs it, locally on your own machine and account.
Runs locally: It runs on your machine and your account instead of a company cloud, using the coding bot you already pay for like Claude Code, Codex, or Cursor.
1,200+ integrations: Each bot gets its own screen, schedule, and tools, and connects across more than 1,200 integrations that every bot shares once linked.
Approval gate on writes: Reads are free, but anything the bot sends out is held until you tap approve, so it never acts without your sign-off.
Try it now → https://github.com/najmuzzaman-mohammad/gawkbot
OpenAI just released ChatGPT Images 2.5, its new state-of-the-art image model, with sharper detail, faster generation, and much better consistency when you edit the same subject across multiple turns. It is rolling out to every ChatGPT tier plus Codex.
Keeps your subjects consistent: It is better at preserving the people and products from your reference photos, so a subject still looks like themselves across a long run of edits.
Up to 50% faster: OpenAI says generation latency drops by up to 50% versus the previous Images 2.0, with more natural lighting and richer textures.
Precise editing: It changes only the element you ask about and leaves the rest intact, even with complex layouts and transparent backgrounds.
Try it now → https://chatgpt.com
Nanobot is a free, open-source, self-hosted personal AI agent framework written in Python. It packs long-term memory, web search, subagents, and scheduled automations into a small, readable core, and runs on your own API key or a local model.
Runs on your terms: It self-hosts on your own machine and works with OpenAI-compatible APIs or local LLMs, so you can use your own key or run everything offline.
Memory, subagents, automation: It keeps long-term memory across sessions, delegates to inline subagents, runs scheduled automations, and can even use your terminal.
Meets you where you chat: It runs in a WebUI, terminal, or chat apps like Telegram, Slack, Discord, and email, with an OpenAI-compatible API built in.
Try it now → https://github.com/HKUDS/nanobot
DeepSeek just open-sourced V4.1-Flash, the smallest model in its new architecture family, with native image understanding and benchmark results the company places ahead of frontier closed models. The weights ship under the MIT license, free to download on day one.
552B MoE, built for efficiency: It is a 552 billion parameter Mixture-of-Experts model using a new causal encoder-decoder design that activates just 8B parameters for input and 16B for output.
Native vision: Unlike text-only releases, it understands images natively, and DeepSeek says it uses roughly a quarter of the KV-cache memory of the model it replaces.
Free and open weights: The full weights are on Hugging Face under the MIT license, so you can download, fine-tune, and self-host with no royalties.
Memanto gives your coding agents permanent memory with a single pip install. It saves your session context, sorts it into typed categories, and pulls the relevant piece back in milliseconds, so your agent stops re-learning your codebase after every reset.
One pip install: Set it up with pip install memanto, with no vector database to provision, no schema, and no backend service to babysit.
13 memory types, sub-90ms recall: It sorts context into 13 typed memory categories like fact, decision, and preference, and retrieves the relevant one in under 90 milliseconds with zero indexing wait.
Works across your tools: It plugs into Claude Code, Cursor, Codex, and 20+ more, and runs on your own API key, a local model, or fully on-prem via Docker.
Try it now → https://github.com/moorcheh-ai/memanto
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