The Top 5 Hugging Face Models of the Week!

Plugged it into an OpenAI compatible API, and wow, the improvement over older Qwens is clear. Qwen announced a new 180B multimodal model which takes both text, images, and videos as inputs with a built-in 262K token context length (expandable to 1M). GPQA Diamond achieved 91.7%, SWE-bench Pro achieved 62.5% — I plugged it into my agent and suddenly tasks that used to stump him are working perfectly. Note that this is still an experimental architecture release, and the license is qwen-community instead of Apache so make sure to check your legal department.

💻 Server-grade hardware — you can't run this at home.

I opened up HuggingChat, a free web-based chat that doesn’t even require an API key! Zhipu AI published the first model in the GLM-5 family, a total of 320B parameters, but only 18B are actually activated at any given time, making it very responsive. It supports images and text, has a 300K token context length, and a parameter to control its reasoning intensity for different tasks. At first I didn’t believe the scores (checked twice), but DeepSWE 63.4? Wow! Also, it’s licensed under MIT so you can easily integrate it into your own products.

💻 Server-grade hardware — you can't run this at home.

I played around with the Kimi API, which gives you 1 million tokens of context so you can throw in an entire codebase and let it do its magic. Moonshot AI just might have published the biggest open weight multimodal model ever: 2.78 trillion parameters, handles text, images, and videos all at once. It’s accessible via an OpenAI compatible API, so your agents don’t need any code changes. Integration with HuggingChat is already available for free web use.

💻 Server-grade hardware — you can't run this at home.

I downloaded it and tried it out on some tasks: create an image based on a description, insert text into an image, change the style of a photo — all with just one model and 8 billion parameters. SenseNova published a multimodal image generation/editing checkpoint which can generate, edit images, and render readable text inside images, all in one step. Licensed under Apache 2.0, so feel free to use it anywhere with no limitations. It’s small enough to run on a laptop with a GPU.

💻 A laptop or PC with a GPU of 8 GB or more.

I tried it out on CPU, and it seems pretty quick (<0.5 GB memory usage, looks good). Audio8 published a small TTS model with zero-shot voice cloning abilities: simply input some text and 3–5 seconds of reference audio, and it will output speech in the desired voice. Supports multiple languages. Free for projects with less than $2M in annual revenue, perfect for most independent developers.

💻 Any computer — runs on CPU, takes under 0.5 GB.