Google opened Gemini 4 Argon to developers and Google AI Pro and Ultra subscribers, Reflection AI released the weights for Beam, a 501 billion parameter coding model that scores about 80% on Terminal-Bench, and DeepSeek's V4.1-Flash, with 552 billion parameters and 16 billion active, outscores GLM-5.3 on Terminal-Bench with weights you can download for free. And that's not all, here's everything in AI this week:
NVIDIA's research lab has released SANA-WM, a 2.6 billion parameter open-source world model. Give it an image, a text prompt, and a camera path, and it generates a minute of 720p video that follows the camera on a single GPU.
Precise camera control: The camera can move in all six directions (forward, sideways, up, and rotation), and the model follows the path you set across the full minute.
Runs on one GPU: A distilled version generates a 60-second 720p clip in 34 seconds on an RTX 5090.
Efficient for its quality: NVIDIA reports visual quality comparable to the closed LingBot-World and HY-WorldPlay systems at 36x higher throughput. The license is Apache 2.0.
Together with Lovable
Turn your expertise into working software
Describe what you want to build
Get real software, not a prototype
Lovable handles payments, logins, databases
Change any detail yourself, any time
Put it in front of customers this week
Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.
PentAGI is an open-source, self-hosted system of AI agents that runs penetration tests on systems you are authorized to test. You give it a target and a goal, and it plans the work, runs the tools, and records what it finds.
A team of agents: An orchestrator agent assigns work to a researcher that looks up known vulnerabilities, a developer that plans attacks, and an executor that runs commands.
Sandboxed tools: It runs more than 20 security tools inside isolated Docker containers and stores findings in memory for later assessments.
Bring your own model: It works with OpenAI, Anthropic, Google Gemini, Ollama, DeepSeek, OpenRouter, and others. The license is MIT, and the repo has 30.6k GitHub stars.
Try it now → https://pentagi.com
Reflection AI has released the weights for Beam, a 501 billion parameter coding and agent model, under the Apache 2.0 license. It scores about 80% on Terminal-Bench 2.1 while using about 25,000 tokens per attempt, compared with 80,000 for GLM-5.2 at a similar score.
Mixture-of-Experts: Only 23 billion of its 501 billion parameters are active per token, and the weights fit on an eight-GPU server in 8-bit or 4-bit.
Coding scores: Reflection reports 80.9 on SWE-Bench Verified and 80.1 on Terminal-Bench 2.1. Kimi K3 and GLM-5.3 still score higher on Terminal-Bench.
Free to start: The weights are on Hugging Face in BF16, FP8, and NVFP4, and a public beta API is free to start. Reflection also open-sourced Mirror CLI, a terminal coding agent trained alongside Beam.
Try it now → https://platform.reflection.ai
Gemini 4 Argon, Google's most capable model, became generally available on October 13. It is built for long coding, agent, legal, and finance tasks, and it can write up to 1 million tokens in a single response.
Where to get it: It is in the paid Gemini API, in the Gemini app and AI Mode for Google AI Pro and Ultra subscribers, and in Antigravity for Ultra subscribers. Free users do not get it.
Benchmark results: In Google's own table, Argon beats GPT-6 Astra on 14 of 19 benchmarks. Astra still leads on FrontierSWE v2, 65.5% to 55.0%.
Introductory price: The API costs $2 per million input tokens and $10 per million output tokens until November 30, then $4 and $20.
Try it now → https://gemini.google.com
Locally Uncensored is a free, open-source desktop app for Windows and Linux that combines chat, image generation, video generation, and a coding agent. Everything runs on your own machine, with no Docker, no terminal, and no config files.
One installer: A setup wizard finds or installs a local AI engine and ComfyUI, and the model manager marks which models fit your hardware.
Four tabs: Chat, Create, Code, and Agent share the same local models. The coding agent shows the diff before applying any change and runs your tests afterward.
Hardware: Chat runs a 3B model on 8 GB of RAM with no GPU, and image generation starts at 6 to 8 GB of VRAM. The license is AGPL-3.0, and there is no macOS desktop build.
Try it now → https://locallyuncensored.com
DeepSeek has released V4.1-Flash, a multimodal Mixture-of-Experts model with 552 billion parameters and a 1 million token context window. DeepSeek reports that it beats its own previous flagship, V4-Pro, on performance, cost, and speed.
Small active size: It activates 8 billion parameters when reading input and 16 billion when writing output.
Benchmark results: In Reflection's comparison table, V4.1-Flash scores 90.6 on Terminal-Bench 2.1, ahead of GLM-5.3 at 88.2.
Free weights: The weights are on Hugging Face. The hosted API costs $0.30 per million input tokens and $1.20 per million output tokens.
Try it now → https://openrouter.ai/deepseek/deepseek-v4.1-flash
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