If you have ever typed "hugging face new models" into a search bar, you already know the problem: Hugging Face publishes new models almost every day, and most of them are noise. This page is the opposite of noise. It is a running list of the open-source AI models actually worth your time, grouped by what they do, with honest notes on where each one shines and where it does not.
We update it as new releases ship. Bookmark it, or check back monthly.
What This Tracker Is (and Isn't)
Hugging Face is the largest open hub for AI models — the "GitHub of machine learning." Labs like Meta, Google, Microsoft, Alibaba (Qwen), and DeepSeek publish trained models there that anyone can download and run. This tracker covers the releases that matter for real work: chatting, coding, making images, and transcribing audio. We skip research demos with no practical use.
Chat & Language Models
Open-source · runs locally or via API ·
Llama (Meta) — The Llama family remains the default starting point for self-hosted chat. Newer versions are strong at following instructions and run on a single consumer GPU at smaller sizes. Best for: private assistants, RAG apps, and anyone who does not want to send data to a closed API.
Qwen (Alibaba) — Often the surprise winner on benchmarks for its size, with especially good multilingual coverage. The smaller "instruct" versions are excellent for non-English users and run on a laptop.
DeepSeek — Known for reasoning-heavy tasks at a fraction of the cost of closed models. The distilled versions are popular for coding and math where you want step-by-step thinking.
Gemma (Google) — Lightweight and permissively licensed, a good fit when you need a small model embedded in an app or on mobile.
Phi (Microsoft) — Tiny models trained on "textbook-quality" data; surprisingly capable for their size, ideal for edge devices.
Code Models
Open-source ·
Several of the chat models above also code well, but if code is your main use, look at the "Coder" variants (for example DeepSeek-Coder and Qwen-Coder). They are fine-tuned on repositories and tend to produce cleaner diffs and fewer hallucinated function names than a general model. Best for: autocomplete, refactoring, and generating boilerplate you would otherwise copy from Stack Overflow.
Image Generators
Open-source ·
The Stable Diffusion family is the backbone of open image generation. Newer forks add better prompt understanding, sharper faces, and built-in upscaling. Best for: thumbnails, mockups, and any workflow where you need to generate or edit images without a subscription. Pair it with an open UI like ComfyUI if you want node-based control.
Audio & Voice
Open-source ·
Whisper (OpenAI, open-weight) is still the standard for speech-to-text and handles many languages. For voice cloning and text-to-speech, several open models now produce natural output. Best for: subtitles, meeting notes, and podcast transcription without per-minute fees.
How to Track New Releases Yourself
You do not need to watch the firehose. On Hugging Face: sort the Models tab by "Recently updated," check the Trending page for fast-rising downloads, and follow the organization pages of the labs you care about (meta-llama, google, microsoft, Qwen, deepseek-ai). When a model climbs the trending list within days of release, that is usually the signal it is actually good — not just hyped.
Where to Start
If you want a private chatbot: start with a Llama or Qwen "instruct" model at 7–14B parameters — small enough to run locally, capable enough for daily use.
If you want the best free reasoning: try a DeepSeek "distill" build; it thinks step by step and costs nothing per call.
If you make images: grab a recent Stable Diffusion fork and an open UI; you will never hit a paywall.
None of these require a paid API. That is the whole point of open-source on Hugging Face — you own the model, the data stays with you, and the only cost is compute.