GitHub radar
Top 5 Hugging Face Models This Week
This week turned out heavy in the literal sense: at the top sits a 763-billion DeepSeek that will not run at home, right next to a 1.7-billion model small enough for a phone. The spread in size is unusually wide, and that by itself is worth noting.
I did not even try to put this one on my own laptop: 763 billion parameters need a server, not a home machine. DeepSeek updated their flagship agent and coding model, a small revision of V4.1-Flash with a real gain in reasoning, coding and image understanding. It reads images as well as text and holds up to one million tokens of context, which is a book of about one and a half thousand pages in one go. The weights are open under the MIT license, so I can rent it from a cloud provider instead of being locked to one closed service.
Why a vibe-coder should care
For agent and coding work this is one of the strongest open models this week, and the open license means I am not tied to one vendor's server.
Server-grade hardware — will not run at home
Open on Hugging FaceI checked this one specifically on an ordinary laptop. It weighs only a few gigabytes, and in the 4-bit version it fits into about one gigabyte of memory. LiquidAI built it for local agents: calling tools, giving structured answers, narrow tasks tied to one project. It speaks 16 languages, including Russian, and holds a context of 256 thousand tokens, enough to drop in a long log file or a piece of documentation at once. The authors themselves say plainly that it is not meant for complex multi-step reasoning or agentic coding, and I agree, its strength lies elsewhere.
Why a vibe-coder should care
For an agent that has to call tools quickly and reliably on my own computer or even a phone, this is a good pick, and it is free.
A regular laptop with 8 GB of RAM or more
Open on Hugging FaceI looked at this release specifically for transcription: NVIDIA put out an open, offline speech model that turns audio into text without sending anything to the cloud. It understands 40 languages, tells speakers apart by voice, adds timestamps, and translates speech from English into 39 languages and back. The word error rate on English is around 4.7 percent, and the transcript comes out with almost no manual correction needed. It handles recordings up to an hour in one pass, and for that it really needs a strong machine, not a laptop with four gigabytes of memory.
Why a vibe-coder should care
If I am transcribing interviews or calls and do not want to send the audio to someone else's cloud, this is a working offline replacement for paid transcription services.
A Mac with 64 GB of memory or more, or a GPU with 48 GB
Open on Hugging FaceI expected this one to be about raw recognition accuracy, but the real point turned out to be document length. Baidu removed the usual page-size limit: ordinary OCR models start losing track of the beginning once a document gets long, while this one, thanks to a new attention mechanism, can take a document of almost any length without me splitting it into chunks by hand. The model itself is small, 3.3 billion parameters, and fits on an ordinary laptop. It works across languages and reads page structure, not only the text.
Why a vibe-coder should care
If I need to recognize a long scan or an entire book in one pass, without splitting it up and without a paid subscription, this is exactly the case for it.
A regular laptop with 8 GB of RAM or more
Open on Hugging FaceI put this one on my own working machine, not a server: 33 billion parameters is the level of one strong computer, not a cloud rig. Nanbeige call it a teacher model, the senior one in the line that will later be used to train smaller versions. It holds a context of up to 512 thousand tokens, enough to drop in a whole large code repository at once, and it is built specifically for agent work: coding, office routine, multi-step tasks. The license is Apache 2.0, so it can be used in a commercial project without asking anyone.
Why a vibe-coder should care
For an agent that has to carry office and coding tasks through to the end on my own hardware, this model is the right size, not a toy and not a monster that needs a server.
A Mac with 32 GB of memory or more, or a GPU with 24 GB
Open on Hugging Face▌ More finds