Top 5 Hugging Face Models This Week

This model is available for download using Ollama. It’s a 4-bit version which is only 17GB, so it easily runs on my 32GB Mac. Qwen published a 27.8B multimodal model that supports text, image, video input with a native 262K context size (can be extended up to 1M). There are significant improvements compared to Qwen3.6-27B (OSWorld-Verified improved from 63.9 to 84.3, SWE-bench Pro improved from 53.5 to 61.7, Terminal Bench improved from 63.4 to 73.0). I appreciate that they introduce a reasoning_effort argument which allows us to fine tune the level of reasoning. Instead of letting the model overthink things, we can now decide how much effort it should put into thinking.

💻 A Mac with 32 GB of RAM or more, or a GPU with 24 GB.

I downloaded LTX-2.5 by Lightricks, which is not only a video creator but also an audio-synchronized video creator using one model. Its main improvement is multishot, which can create several related scenes at once while keeping the same character’s appearance in all shots. This was impossible in the previous version. They improved their VAE decoder, so the images are clearer, especially the faces and texts. Also, they changed their text encoder to a custom Gemma 4 12B model, which can correctly process complicated multi-person prompts. They released their weights under a commercial license, allowing you to use them for projects with revenues up to $10M.

💻 A GPU with 24 GB of VRAM, or a Mac with 32 GB of RAM or more (quantization is supported for smaller setups).

MiniMax Music3 — 2.4B parameter model capable of generating complete songs (up to 5 minutes long) including lyrics, instrumentation, and structure (verses, choruses, bridges, outros). Input is a set of verses/choruses and a description of the desired song (genre, BPM, instruments, mood), the model will create a story around them. This model uses a combination of global (8B) LLM for high-level planning and local (0.6B) LLM for low level details, alongside Flow Matching to produce 32 kHz stereo WAV. However, this model doesn’t seem to have an explicit license, which means you probably shouldn’t use it commercially.

💻 A regular laptop with 8 GB of RAM or more.

Deploy MiniMax H3 — a 33B transformer by MiniMaxAI which converts text/images to videos with matching stereo audio upto 2K resolution. This model has been divided into 3 parts: context pre-processing, 768p base generation & 2K upscaling(H3-Regenerate-2K is currently closed-sourced). Used this model on image+scene description tasks, got ~4–15 sec of videos @24fps. This is the first open-weighted model that I’ve come across which supports stereo audio natively(no need of any audio processing pipeline)

💻 A Mac with 32 GB of RAM or more, or a GPU with 24 GB.

Run DeepSeek-V4-Flash-0731 using vLLM on a server environment. A 304B MoE model released under the MIT license by DeepSeek. ‘Flash’ refers to DSpark speculative decoding to achieve fast inferencing while only activating a fraction of the total parameters. On Terminal Bench 2.1, it scored 82.7 compared to DeepSeek-V4-Pro Preview’s 72.1. On DeepSWE, it increased from 12.8 to 54.4 and on AutomationBench from 12.8 to 25.1. Low/high/max reasoning_effort settings allow you to control how deep/fast you want the model to reason about your task.

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