GitHub radar
Top 3 Hugging Face Models This Week
Week 38 on Hugging Face: DeepSeek's new multimodal Flash with 1M-token context, a 2.5B offline agent that fits on any laptop, and Nex-AGI's model built specifically for long multi-step tasks — not benchmarks.
DeepSeek released V4.1-Flash — a multimodal flagship with up to one million tokens of context. Tested it via API: faster than previous versions, understands images and screenshots right in the conversation. The reasoning mode now scales from 1 to 100 — you set how much thinking to invest per request. MIT license means you can integrate it anywhere without restrictions.
Why a vibe-coder should care
When a task means pushing hundreds of pages of text or an entire codebase through AI in one request — a million tokens changes the situation. Adjustable reasoning saves money: skip the thinking on simple queries, max it out when you need depth.
Серверное железо — дома не запустить. Доступен через API DeepSeek и провайдеры вроде Novita.
Open on Hugging FaceInstalled MiniCPM5-2B from OpenBMB through Ollama — takes 1.5 GB quantized, runs on a laptop without a GPU. A 128K-token context window at this size is unusual: you can feed entire documents in one shot. Tool calling works out of the box: give it a task, the model decides what to run. Ran it on real tasks — solid for 2.5B parameters.
Why a vibe-coder should care
A full AI agent with tool calling and large context — on any laptop, offline, no subscriptions. For when your data needs to stay off someone else's servers.
Обычный ноутбук от 8 ГБ оперативки.
Open on Hugging FaceNex-AGI released Nex-N2.5-Pro — an agentic model built specifically for long, multi-step tasks. Tried it via OpenRouter: hand it a chain of dozens of steps and it holds the thread all the way to the result without losing context. Open weights, Apache 2.0, but too large to run at home. What I liked: the authors are upfront about the goal — real tasks in real environments, not lab benchmarks.
Why a vibe-coder should care
When an agent needs to run a project from first step to final result without losing focus or asking again — this is what it was built for. You can try it through OpenRouter right now without installing anything.
Серверное железо — дома не запустить. Доступна через OpenRouter без установки.
Open on Hugging Face▌ More finds