GitHub radar
Top 5 Hugging Face Models This Week
This week on Hugging Face: DeepSeek's first vision model, Qwen's strongest open release yet that runs locally, a 4B model with 1M context that fits on any laptop, an open local video generator with synchronized audio, and a multimodal model approaching Claude Opus at a fraction of the price.
DeepSeek just added vision to their V4 Flash model — I tested it through their API and it genuinely understands screenshots, charts, and images alongside text. What surprised me most is how well it handles agentic tasks with visual input: sent it a screenshot of a broken UI and it pinpointed the problem without me describing anything. The benchmark numbers on agentic tasks are strong. You can't run 305B locally, but it's available through DeepSeek's API right now.
Why a vibe-coder should care
If you use DeepSeek for agentic work, now you can feed it screenshots, diagrams, and visual context — no more describing what you see in words.
Server-grade hardware — you can't run this at home; try it via the DeepSeek API.
Open on Hugging FaceI've been running Qwen3.8-27B through Ollama all week — Qwen call it their most capable open model to date, and the numbers back it up: SWE-bench Pro at 61.7% and OSWorld-Verified at 84.3%, which means it can actually navigate your computer autonomously. Context goes to 262K tokens natively, extendable to 1M. The quantized version fits in 32 GB of Mac RAM. What I personally liked: you can dial thinking depth down when you don't need full reasoning — much faster for everyday tasks.
Why a vibe-coder should care
The best open-weight agentic model you can run locally right now — coding, computer control, and vision in one package, Apache 2.0 license.
A Mac with 32 GB of RAM or more, or a GPU with 24 GB.
Open on Hugging FaceDownloaded Spark-X2.5-4B from XHToken (SparkLLM Team) and was genuinely surprised: a 4B model that runs on any laptop with 8 GB RAM, yet supports coding, translation, tool use, and structured generation. The 1 million token context window at this size is unusual — most 4B models top out way lower. Over 200 languages. Set it up in LM Studio in about five minutes with no friction.
Why a vibe-coder should care
If your machine can't run a 27B+ model, this is the go-to: a genuinely capable assistant on 8 GB RAM with 1M context, free under Apache 2.0.
A regular laptop with 8 GB of RAM or more.
Open on Hugging FaceLightricks released LTX-2.5 this week — an open video model you run locally from text, images, or existing clips, no cloud subscription needed. What sets it apart: multishot support keeps your character looking consistent across multiple scenes, and audio generates in sync with the video, not pasted on afterward. I hadn't seen an open local model pull off both in one go before. You need a CUDA-capable GPU to run it — this isn't a laptop project.
Why a vibe-coder should care
Generate video with synchronized audio locally, consistent characters across scenes — no monthly subscription, open weights you own.
A CUDA-capable GPU with 24 GB of VRAM or more.
Open on Hugging FaceGLM-5.3-Flash from Zhipu (Z.ai) is their first natively multimodal model in the GLM-5 series — 321B total parameters but only 18B active at inference time (MoE architecture), which makes the API cost much lower than the parameter count suggests. The team claims it approaches Claude Opus 4.8 on coding and agentic benchmarks at one-tenth the price. I plugged it into an agent pipeline via the Z.ai API as a cheaper replacement for heavy reasoning steps. Context is 300K. English and Chinese only — not a universal language model.
Why a vibe-coder should care
Multimodal reasoning model at a fraction of Claude Opus price — plug it into your agent pipeline via API, MIT licensed.
Server-grade hardware — you can't run this at home; use the Z.ai API instead.
Open on Hugging Face▌ More finds