• o3-mini-level results, but open-weight under Apache-2.0, so you can run it locally (and including for commercial purposes).
• Mixture-of-Experts design → activates only a fraction of params per token; 128k context.
• Runs on ~16 GB VRAM / laptop-class hardware; downloads on Hugging Face.
• Tool use & agent-style workflows (browse/code/execute) are supported.
• Caveats: strongest on text/STEM; multilingual & long-horizon reliability still TBD. Probably worse than other models since training was focused on English.
Verdict: real progress in small-model efficiency + reasoning. Worth trying if you want local/edge deployments (and in english!), just don’t expect closed-model polish on day one.
—
Louis-François — PhD-dropout & CTO/co-founder @ Towards AI.
Follow for tomorrow’s no-BS roundup.
#OpenAI #GPTOSS #LLM #MoE #EdgeAI #AInews #HypeOrNo
source