AI NewsOctober 9, 2026

d1-3B and d1-omni-600M don’t write any text at all. Given a question with a fixed set of possible answers, they return one in a single step, in as little as 8 milliseconds, making them far cheaper than a model that has to generate a written response.
Liquid AI released both models as open weights on October 7. d1-3B handles text and images, while d1-omni-600M, a smaller 587-million-parameter early research release, also handles text paired with audio. On Liquid’s own Decision Index benchmark, d1-3B scored 48.57, which the company says matches a decision model 12 times its size.
These are meant for jobs like sorting support tickets, routing requests, or flagging spam, where a short, structured answer is all that’s needed, not for open-ended conversation. Both run under a license that allows free commercial use for companies under $10 million in annual revenue.
Key Capabilities:

• No text generation: Returns a direct answer in one pass, not written word by word.
• Fast: As little as 8 milliseconds per answer on a consumer-grade GPU.
• Free below $10M revenue: Commercial use is free for companies under that revenue threshold.

Mellum2.1 keeps the same architecture as its predecessor but gained the ability to explore a repository, edit files, and check its own work, after a summer spent training it in real sandboxed coding environments.
JetBrains released Mellum2.1 on October 8, a 12-billion-parameter model that only “activates” 2.5 billion of those parameters per request, keeping it fast enough for routing and quick sub-agent tasks. The upgrade over June’s Mellum2 comes almost entirely from reinforcement learning rather than a different architecture, and JetBrains reports roughly 1.6x faster responses thanks to a new technique that predicts multiple words at once.
It’s released under the Apache 2.0 license, meaning it’s free to use commercially, and is aimed at developers who want a small, self-hostable coding agent rather than relying on a larger cloud model for every task.
Key Capabilities:

• Same size, smarter: Unchanged 12B/2.5B architecture, but trained through RL to self-verify its edits.
• Faster responses: Roughly 1.6x faster per request via multi-token prediction.
• Free to self-host: Apache 2.0 license, available now on Hugging Face.

Most search systems scan a document as text, losing tables and layout. Perplexity’s new models instead treat a page as an image, so they “see” a PDF, slide, or scan the way a person would, and a small model can search an index built by a much bigger one.
Perplexity released two open-weight models on October 7: a 0.6B version built for phones or edge devices, and a 9B version built for building a high-quality search index. Both convert text, images, and page images into the same shared numerical format, so a developer can build a large, accurate index with the 9B model, then run fast day-to-day searches with the small one. On MADQA, a test of answering questions from PDF documents, the 9B model scored 92.4%, the best result among retrieval-only systems tested, though it still trailed a more complex agentic search system.
Both are free for commercial use under the MIT license, available now on Hugging Face, with a hosted API still to come.
Key Capabilities:

• Mix and match: A 9B-built index can be searched using the smaller, cheaper 0.6B model.
• No OCR needed: Reads PDFs and slides as images directly, keeping tables and layout intact.
• Top score: 92.4% on MADQA, the best among pure retrieval systems Perplexity tested

Instead of switching between separate tools for email, docs, and code, Google’s new Gemini agent takes one plain-language goal and works across a company’s existing apps to deliver finished output.
Google Cloud unveiled the Gemini agent on October 8 at its Gemini at Work event, held at NASA’s Hangar One. It runs persistently in the cloud, so it keeps the same memory and context whether accessed from Gmail, Slack, or Microsoft 365, and it can spin up specialized “coworker agents” with their own identities for recurring roles. It also picks the best underlying model for each task, reportedly including Google’s own Gemini 4 Argon alongside Anthropic’s Claude Opus 5.5 and Sonnet 5.5.
Google said nearly 500 Cloud customers now process over 1 trillion tokens each per year, and about 90% of Fortune 100 companies use Gemini Enterprise. No separate pricing or full rollout date has been announced yet.
Key Capabilities:

• Persistent memory: Runs in the cloud, keeping one shared context across every app and device.
• Sub-agents: Can spawn specialized coworker agents with defined roles for recurring tasks.
• Model flexibility: Reportedly supports pulling in outside models, including Anthropic’s Claude, alongside Google’s own.

Odyssey-3 is a “world model,” a type of AI trained to predict how a scene will change next, including how actions affect outcomes. One version just posted the highest known score on a physics-accuracy benchmark.
Odyssey launched Odyssey-3 on October 8 as a free research preview, open to the public. The company says the same underlying model can control robot arms and humanoids, drive a real car, fly a drone, and generate playable game environments from a text prompt, all without being separately trained for each task.
Odyssey-3 Pro scored 66.1 on Physics-IQ Verified’s video-to-video test, built by Anates Labs and DeepMind to check whether a model correctly continues real physics experiments like fluid motion or magnetism. Odyssey calls that the highest reported score on the leaderboard, beating entries from Nvidia and Black Forest Labs. Odyssey raised $310 million in June at a $1.45 billion valuation.
Key Capabilities:

• One model, many bodies: The same system reportedly drives cars, flies drones, and runs humanoid robots.
• Benchmark claim: 66.1 on Physics-IQ Verified video-to-video, the highest reported score as of launch.
• Free preview: Open to try now at no cost through Odyssey’s research preview.

Ask it a plain-English question about company data and Claude Dashboards returns a chart that updates as the underlying data changes, while a separate tool, Motion, turns reports into short editable animations.
Anthropic launched Claude Dashboards and Claude Motion in beta on October 8. Dashboards connects to platforms like BigQuery, Snowflake, Databricks, and Salesforce, and can hand off results to tools like Grafana or Mixpanel for deeper digging. Motion writes the animation as editable code rather than using a video generation model, so Anthropic says there’s no invented footage or AI-generated people, and every word or timing can be adjusted afterward.
In the same announcement, Claude Docs, Slides, and Design moved out of beta entirely, now free for everyone including the Free plan. Anthropic says over 45 million documents, decks, and designs have been made in Claude so far.
Key Capabilities:

• Live data link: Dashboards connects directly to a company’s data warehouse, not static uploads.
• Editable animation: Motion outputs are code, so every element can be changed before export.
• Access split: Dashboards is on all paid plans; Motion is Team and Enterprise only.