Occamy-1.0 targets long-horizon agent work at 35B

A new open model called Occamy-1.0 is making the case that mid-sized systems can hold their own on the kinds of drawn-out, multi-step tasks that increasingly define practical AI work. According to the accompanying paper on Hugging Face, the model is a 35-billion-parameter dense system positioned on the Pareto frontier of capability versus size, and it is aimed squarely at "co-work" scenarios where a model acts as an ongoing collaborator rather than a one-shot oracle.

The framing matters because most benchmarks reward quick, single-turn answers, while real agent deployments demand something harder: sustaining a plan across many steps, recovering from mistakes, and staying coherent over long horizons. Occamy-1.0 is explicitly tuned for that pattern, with reasoning listed among its core modalities alongside text generation.

Why it matters

At 35B parameters, Occamy sits in a practical middle ground — large enough to reason competently, small enough to run without the infrastructure that frontier-scale systems require. That makes it a candidate for teams that want capable agents they can host themselves.

• Dense 35B architecture, not a mixture-of-experts design
• Focused on long-horizon, multi-step and reasoning tasks
• Released as an open model for co-working use cases

Some details remain unspecified in the record, including context length and exact licensing terms, which will shape how broadly the model can be adopted. Still, as the initial release of a new family, Occamy-1.0 signals continued momentum around open models built for agentic work rather than leaderboard sprints. The full technical claims are laid out in the Hugging Face paper.