Top AI Product

Korea’s Upstage just put out Solar Open 2, an open-weight LLM rebuilt from scratch for agentic work — the multi-step, tool-calling, document-heavy stuff that eats context. It’s a Mixture-of-Experts: 250B total parameters, but only 15B activate per token. They reused just 2.3% of the last generation. This is a ground-up model wearing an old name.

The trick is the attention stack. Three linear-attention layers for every one softmax layer, so a 1M-token context runs at roughly a quarter of the VRAM a full-softmax design would need. Quantized, it fits on two H200s. Upstage claims it beats DeepSeek V4 Flash and Mistral Medium 3.5 on agent benchmarks — frontier scores off a 15B active budget.

Where you can run it

Weights are live on Hugging Face under an Apache-2.0-based license, commercial use allowed. Self-host it, or grab the NVFP4 quant if you’re tight on GPU. Point your agent framework at it and go.

Why it matters

This is a national project. Backed by Korea’s sovereign-AI initiative, Upstage plans to ship it into Daum — 10M weekly users — and build local NPU infra for finance, legal, medical, and government. A new open-weights player, and it’s state-funded.

You Might Also Like

• Ggml Llama cpp Joins Hugging Face and Honestly it was Only a Matter of Time
• Pollen Robotics Reachy Mini a 299 Desktop Humanoid That Runs 1 7m Hugging Face Models
• Title x Square Robot Wall a With Wall oss 276m From Xiaomi 62 dof Weights on Hugging Face
• Warp Open Source Agentic Terminal Hits 50k Stars in a Week
• Tencent Hunyuan hy3 Preview Goes Open Source 295b moe 21b Active 256k Context