Technology
November 15, 2026 1009 words

The Open-Weight AI Wave: LongCat-2.5, ERNIE 5.5 Flash, Ripple-65B and More

Hams AI Tech
Professional Content Team
Expert in AI Prompts & Professional Tools

If 2025 was the year open models caught up, late 2026 is the year they flooded the market. In the first half of November alone, labs across China, the United States, Europe, Korea and Japan published weights for models ranging from a few billion to nearly three trillion parameters. This roundup covers the most important releases, explains what their licenses allow and shows how professionals can benefit even if they never download a model file.

LongCat-2.5 from Meituan

Meituan released LongCat-2.5 on November 2 and published its weights a week later under the MIT license, which allows commercial use, modification and redistribution. It is a natively multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token. It targets long-horizon agentic coding across whole codebases, visual coding, creative design and 3D tasks. Meituan's benchmarks place it between Claude Opus 4.8 and Claude Opus 5 on most tests, and it was trained entirely on Chinese domestic accelerators. A quantized W4A4 checkpoint is also available, and the LongCat API works with Claude Code, Codex and Hermes harnesses.

ERNIE 5.5 Flash from Baidu

Baidu released ERNIE-5.5-Flash-Base on November 5, a 280 billion parameter sparse mixture of experts with 14 billion active, trained on 33.5 trillion tokens. This is a base checkpoint with no post-training. On November 11, Baidu followed with ERNIE 5.5 Flash Preview, its first post-trained model on that base, offering a one million token context and more than 200 tokens per second. Baidu says the preview weights will be open-sourced under Apache 2.0.

Ripple-65B Preview from Ricursive

Ricursive Intelligence released Ripple-65B Preview on November 4, a 65 billion parameter dense model pretrained from scratch on 12 trillion tokens with up to one million tokens of context, aimed at agentic coding. Ricursive reports a US open-weight state of the art on Terminal-Bench 2.1 at 80.52, ahead of much larger open models, at a full research and development cost of about 32 million dollars. Weights are under Apache 2.0, with API access through Together AI, GMI Cloud and OpenRouter.

Hydrogen-1.0 from Deep Principle

Deep Principle, a Hangzhou AI-for-science company, released Hydrogen-1.0 on November 2, a 753 billion parameter model for chemistry, materials science and life sciences, post-trained from Z.ai's GLM-5.3. It is designed to help researchers reason through scientific questions, write simulation code, plan experiments and interpret results such as NMR spectra. Deep Principle reports 93.0% on GPQA Diamond. Weights are MIT licensed. Scientists can explore our science AI resources.

Edge0-145B and Edge0-120B

Edge0 released open-weight models designed for laptops on November 12. Edge0-145B is a sparse mixture of experts with about 10 billion parameters active per step, reporting 29 gigabytes of peak memory and more than 80 tokens per second on a compatible laptop with Apple M-series or NVIDIA RTX hardware. The license is Edge0's own community license, not a standard open-source license. Our article on running AI locally covers this category in depth.

Other Notable Releases

    • SOV-Transcribe-1 from Salesforce, a 26 billion parameter speech recognition model under Apache 2.0.
    • Arete-27B from Nous Research, a fine-tune of Qwen3.8-27B that reduces propaganda framing on politically sensitive topics.
    • PLaMo 3 Translate from Preferred Networks, translation models in four sizes for 53 languages.
    • ZONOS3-Mini from Zyphra, a compact voice-cloning text-to-speech model.

Understanding Open-Weight Licenses

Open weight does not always mean you can do anything you like. Licenses fall into three broad groups:

    • Permissive: MIT and Apache 2.0, used by LongCat-2.5, Ripple-65B, Hydrogen-1.0 and Inkling 2 Small. Commercial use is generally allowed.
    • Modified permissive: such as Mistral's modified MIT license, which adds conditions for model-as-a-service businesses.
    • Custom community licenses: such as those for Kimi K3.1 Pro, Edge0 and MiniMax H3 models, which may restrict scale, territory or use cases.

Before building a product or client service on an open model, have counsel review the license. Technology lawyers can speed up that review with tools like Claude Skills for Business Lawyers and our contract law resources.

Why Professionals Should Care

Even if you never host a model, the open-weight wave benefits you:

    • Lower prices across the market, as proprietary providers compete with free alternatives.
    • More private deployment options from vendors who host open models in dedicated environments.
    • Specialized models for science, translation, speech and coding that would not exist otherwise.
    • Greater bargaining power, because switching providers is easier when good alternatives exist.

How to Take Advantage

    • Ask your software vendors which models power their tools and whether private hosting is available.
    • For sensitive work, consider providers that run open models in your region.
    • Build your prompts and workflows in a model-agnostic way so you can switch when better options appear.

Frequently Asked Questions

Which open model is best for coding?

It depends on your hardware and budget. Ripple-65B targets agentic coding at a manageable size, while LongCat-2.5 and Kimi K3.1 Pro are much larger. Test with your own codebase.

Can I run these models on a normal computer?

Most cannot. Laptop-focused models such as Edge0-145B are the exception. Large models need GPU servers or hosted APIs.

Are open models safe to use with client data?

They can be safer than public APIs when hosted in a controlled environment, but safety depends on how you deploy, secure and monitor them.

The Bottom Line

The open-weight wave is reshaping the AI market, lowering costs and expanding choice. Professionals who keep their workflows portable will benefit most. Explore our agentic AI resources and programming tools to build workflows that work with any model.

Related AI News

Topics & Keywords
open weight models open source AI LongCat-2.5 ERNIE 5.5 Flash Ripple-65B Hydrogen-1.0 Hugging Face