Ali Asif

The AI log, frontier · open weights · image, video, world · agents and coding

Covers 15 Dec 2025 to 21 Dec 2025

Gemini 3 Flash, GPT-5.2-Codex and Nvidia's open Nemotron 3

Week of 15 Dec 2025: Google makes Flash the default, OpenAI ships a Codex model for long coding jobs, and Nvidia opens a small agent-ready model family.

By , published

What shipped

  1. Google

    Gemini 3 Flash

    Google put its fast tier of Gemini 3 into the Gemini app and AI Mode in Search as the default, so most users now get a frontier-class model without picking one.

  2. OpenAI

    GPT-5.2-Codex

    OpenAI's agentic coding model, built for long professional software jobs and defensive security work, landed about a week after GPT-5.2.

  3. NVIDIA

    NVIDIA Nemotron 3 Nano 30B-A3B

    An open mixture-of-experts model with only about 3B active parameters, aimed at cheap multi-agent workloads.

  4. Xiaomi

    MiMo-V2-Flash

    Xiaomi entered the open-weights race with a fast mixture-of-experts model built for reasoning and agent use.

  5. Alibaba Qwen

    Qwen-Image-Layered

    It splits one image into separate transparent layers, which makes real editing possible instead of re-rolling the whole picture.

  6. Meta

    SAM Audio

    Meta's segment-anything idea moves to sound: you describe or point at a source and the model pulls it out of a mix.

  7. Google

    FunctionGemma 270M

    A tiny Gemma tuned for function calling, built to run on a phone or a laptop and route requests to tools.

This was the last full week before the holidays, and the labs clearly knew it. A lot of launches got squeezed in, and the theme was speed and price more than raw ceiling.

Gemini 3 Flash

Google's headline move was making Gemini 3 Flash the default model in the Gemini app and in AI Mode in Search. That is a distribution decision as much as a model decision. A very large number of people will now hit a Gemini 3 class model without ever choosing it.

The reason I care is the tier, not the brand. Fast models used to be the ones you tolerated. Here the fast model is the one Google is betting its front door on, which tells me the quality gap between "fast" and "best" is shrinking for everyday work.

For anyone running agents all day, that changes the maths. Most of what my agents do is dull and repetitive, and it never needed the most expensive model in the first place.

Coding agents

OpenAI followed GPT-5.2 with GPT-5.2-Codex, tuned for long agentic coding sessions and defensive security. I read the release as a signal that the coding lane is now its own product line, separate from the chat model.

Open weights

Nvidia released Nemotron 3 Nano, a 30B mixture-of-experts model with roughly 3B parameters active at a time. It is built for running many small agents cheaply. Xiaomi also put out MiMo-V2-Flash, a fast open model aimed at reasoning and agent use. Google added FunctionGemma, a 270M model that does one job: turn a request into a tool call.

Image and audio

Alibaba's Qwen-Image-Layered decomposes an image into RGBA layers you can edit separately. Meta's SAM Audio brings prompt-based segmentation to sound, so you can isolate a voice or an instrument from a mix.

A note on how I use these releases. I do not switch models the day something launches. I keep a short list of real tasks, such as reading a supplier quote and flagging missing line items, and I rerun them when a new model arrives. If the result is better or cheaper, it goes into the workflow. If not, I read the announcement, file it, and move on. It sounds dull, but it keeps me from rebuilding my setup every Thursday.

What I'd watch next: whether the open models from Nvidia and Xiaomi hold up on real agent tasks once people run them outside the vendor demos.

Sources

Topics: OpenAI, Qwen, Alibaba Qwen, Codex, Gemini, Google, NVIDIA, Meta