Weekly · 6 min read
- Meta
- NVIDIA
Meta releases AI models emphasizing memory and extended tasks
October 12 (Mon) - October 18 (Sun), 2026
Summary
AI news for October 12–18: Meta's Muse Wonder and its memory for long-running work, Gemini 4 Argon's wider access and API rates that double in December, NVIDIA's local models, and free document checks with Google's Deep Review.
From October 12 to 18, Meta released new models built around memory and long-running work, and Google opened Gemini 4 Argon to more people through its app, Search and API. NVIDIA announced models you can run on your own hardware, and Google Research made Deep Review, a tool that checks a document's reasoning, available to everyone.
Put the four side by side and a pattern shows. AI is taking on more multi-step work, and you have more say over where that work runs and what it costs.
| Date | Announcement | What changed |
|---|---|---|
| Oct 12 · Meta | Muse Wonder and Full launch | Memory across conversations and sustained work |
| Oct 13 · Google | Gemini 4 Argon access expands | App, Search and API access; launch pricing until the end of November |
| Oct 14 · Google Research | Deep Review opens to everyone | Reasoning checks for PDF and text, two free reviews a week |
| Oct 15 · NVIDIA | Nemotron 3.5 Super VL and Bolt | A model that reads documents and models you run yourself |
01 · Muse Wonder
Memory for work that continues
Meta launched Muse Wonder and Muse Wonder Full on October 12. It highlights two features. Active Memory pulls context from earlier conversations when it's relevant to the current task, and Active Focus keeps working toward a goal by trying other approaches when it gets stuck. Meta announcement
| Model | Where you can use it | API price per million tokens |
|---|---|---|
| Muse Wonder | Meta Model API and Muse Code | $4 input / $20 output |
| Muse Wonder Full | Supported Muse Code subscriptions | No separate API price |
Access and pricing are as listed on Meta's model page on October 19.
The cat asks Q&A
The researcherWhat makes it an agent model rather than an AI that answers questions?
Pocket NotebookAn agent takes a goal and works through several steps on its own, often using tools. Instead of "explain this code," you hand over "fix it, test it, and fix it again if it fails."
- The researcher
Is Active Memory just saved chat history?
Pocket NotebookIt goes further than storing it: Meta says it picks out what the current task needs from past conversations. The point is to stop you re-explaining a project's requirements every time.
- The researcher
Can I use Full through the API too?
Pocket NotebookIt's offered differently. Full comes with Muse Code subscriptions, and Wonder is the one with published API pricing.
02 · Gemini 4 Argon
Wider access and API pricing
Google expanded access to Gemini 4 Argon on October 13. It was first introduced on September 30, and it's now available to Pro and Ultra users in the Gemini app and Search AI Mode, and to developers through the API. Google announcement
The API comes with launch pricing that has an end date. These are Google Cloud's announced API rates, separate from app subscriptions.
| API usage date | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Through November 30, 2026 | $2 | $10 |
| From December 1, 2026 | $4 | $20 |
The same usage costs twice as much after the switch. Here's the math for one million uncached input tokens and 100,000 output tokens.
| Usage date | 1M input tokens | 100K output tokens | Total |
|---|---|---|---|
| Through November 30 | $2 | $1 | $3 |
| From December 1 | $4 | $2 | $6 |
The cat asks Q&A
- The researcher
What's an API, exactly?
Pocket NotebookIt's how your program sends a request to the AI and gets an answer back. Instead of chatting in the app, you can build the model into your own service.
- The researcher
And the tokens in "per million tokens"?
Pocket NotebookThey're the pieces of text a model processes. A token isn't always one character, and the text you send (input) and the answer you get (output) are billed separately.
03 · NVIDIA Nemotron 3.5
Document and agent models on your own hardware
NVIDIA announced Nemotron 3.5 Super VL and the Bolt model family on October 15. Super VL reads the text, tables and charts in documents, and Bolt focuses on running agent tasks directly on supported NVIDIA hardware. NVIDIA announcement
| Model | Role | Status as of October 19 |
|---|---|---|
| Super VL | Visual information such as documents, tables and charts | Available |
| Bolt 74B-A9B 24GB | Local tasks on supported RTX and DGX Spark systems | Available |
| Bolt E4B 4GB | Small conversational and agent use | Scheduled for October 23 |
The cat asks Q&A
- The researcher
Does running locally mean no internet needed?
Pocket NotebookIt means the model's computation happens on your PC or hardware you manage. Features like web search may still need a connection.
- The researcher
What are "open weights"?
Pocket NotebookFiles containing a trained model's numbers that you can download. Instead of calling an external API, you can run and manage the model yourself.
04 · Deep Review
An AI that checks a document's claims and evidence
Google Research made Deep Review available to everyone on October 14. It began as a tool for reviewing research papers and now covers other kinds of documents. It currently accepts PDF and plain text, and each user gets two free reviews a week. Official announcement
According to Google, it doesn't read the whole document in one pass. It splits the document into sections, allocates review effort to each, has several AI tasks review them in parallel, then combines the results and checks the document's consistency as a whole.
%% How Deep Review checks a document · simplified from Google Research's description
flowchart TD
A["Upload · PDF or text"] --> B["Split into sections · allocate review effort"]
B --> C1["Review section 1"]
B --> C2["Review section 2"]
B --> C3["Review section 3"]
C1 --> D["Combine results · check overall consistency"]
C2 --> D
C3 --> D
D --> E["A person reads the findings and decides"]
The cat asks Q&A
- The researcher
How is that different from "summarize this document"?
Pocket NotebookA summary makes the content shorter. Deep Review checks whether the claims and the evidence hold together, like a missing assumption in a design doc, or a slide whose conclusion doesn't match its chart.
- The researcher
What happens to the documents I upload?
Pocket NotebookGoogle says they aren't used for training and are kept for one year so the results can be shown in the interface.
To close
What to take away — Matching the news to your situation
Here's how this week's news lines up with different situations.
| Your situation | What to look at |
|---|---|
| You keep handing the same project to an AI | Muse Wonder's memory and sustained work |
| You want AI inside your own service | Argon's access routes and API costs from December |
| You want to run a model on your own hardware | Bolt's availability, supported hardware and running costs |
| You want to check a document's reasoning | Deep Review's free reviews, compared against the original |
These four stories were chosen on editorial grounds, not as a popularity ranking. Product details and prices come from the original sources as of October 19; we haven't used the products ourselves or benchmarked them.
Sources
- Meta announcementresearch.meta.ai
- Meta's model pagedev.meta.ai
- Google announcementblog.google
- Google Cloud's announced API ratescloud.google.com
- NVIDIA announcementblogs.nvidia.com
- Official announcementresearch.google
- Deep Reviewresearch.google