Ali Asif

The AI log, frontier · open weights · image, video, world · science

Covers 3 Aug 2026 to 9 Aug 2026

Qwen3.8-Max at 2.4T, Motif 3, and WeatherNext goes open

Week of 3 August 2026: Alibaba ships the 2.4T Qwen3.8-Max, Motif Technologies opens a 314B model, and DeepMind publishes cyclone forecasting in Nature.

By , published

What shipped

  1. Alibaba

    Qwen3.8-Max

    Alibaba's flagship is a 2.4-trillion-parameter mixture-of-experts model with 95B active parameters and a 1M-token window, pitched at multi-day coding projects.

  2. Motif Technologies

    Motif 3

    A 314B mixture-of-experts model with 13.2B active parameters, MIT license, trained from scratch on about 12.5T tokens for agentic and terminal tasks.

  3. Alibaba

    Qwen-Image-3.0

    The standard tier of the new image series: faster and cheaper than the Pro model, with the same generation and editing interface.

  4. Google DeepMind

    WeatherNext cyclone forecasting

    A Nature paper reports three-day cyclone forecasts as accurate as older models' two-day ones, and DeepMind is open-sourcing the models.

  5. Liquid AI

    LFM2.5-2.6B

    A 2.69B on-device reasoning model with 131K context and 16 languages, meant for agents and data extraction on phones and laptops.

  6. Cactus Compute

    Needle 2

    A 45M-parameter open model in roughly 14 MB that does tool calling and structured extraction on a device.

  7. arXiv

    Recursive Synthesis for Long-Horizon Terminal Tasks

    Tackles the cost of training data for terminal agents, which can run to hundreds or thousands of dollars per task.

Alibaba went very big, a Korean startup went open, and DeepMind put out the kind of science result that has a use outside a lab.

Qwen3.8-Max

Alibaba released Qwen3.8-Max on 3 August. It is a 2.4-trillion-parameter mixture-of-experts model with about 95B parameters active per token, a 1M-token context and up to 128K tokens of output. Alibaba pitches it at coding and professional work, with claims about delivering complete projects that run for ten days or more and about native visual understanding across planning, execution and checking. Those are vendor claims and I'd test them on a real job before believing the ten-day figure. There is a Hugging Face page under a custom Qwen3.8-Max license, so read the terms.

A related image release followed on 4 August: Qwen-Image-3.0, the faster and cheaper standard tier next to last month's Pro model.

Motif 3

Motif Technologies released Motif 3 on 7 August. It is a 314B mixture-of-experts model with 13.2B active parameters, trained from scratch on about 12.5 trillion tokens with extra Korean, reasoning and agentic data. It has a 262,144-token native context, a built-in multi-token prediction head for faster decoding, and an MIT license. For a lab outside the biggest names, shipping from-scratch weights this size under MIT is a strong signal.

WeatherNext, in Nature

On 6 August Google DeepMind reported in Nature that its WeatherNext model predicts a tropical cyclone's path, strength and wind structure better than existing methods. Its three-day forecasts match what earlier models reached only at two days, which the company describes as roughly a decade of progress in meteorology. DeepMind is open-sourcing WeatherNext Cyclones and WeatherNext 2, plus a smaller WeatherNext 2-mini.

Small and on-device

Liquid AI released LFM2.5-2.6B on 4 August, a 2.69B reasoning model with 131,072-token context and 16 languages. Liquid does not recommend it for agentic coding, and says so plainly, which I appreciate. Cactus Compute's Needle 2 followed on 9 August: 45M parameters in about 14 MB, built for tool calls and extraction.

One paper worth a look: Recursive Synthesis for Long-Horizon Terminal Tasks, which goes after the cost of building training tasks for terminal agents.

What I'd watch next

Who builds on the open WeatherNext weights first, and whether Qwen3.8-Max gets an open sibling.

Sources

Topics: Google DeepMind, arXiv, Qwen, Liquid AI, Alibaba, Cactus Compute, Weathernext