Skip to content
Predictive Systems
PSI Daily

Hosting Open-Weight Models: Alibaba’s Occamy Agent as a Case Study

SovereigntyAllan C. Tan, MS

Occamy is a 35B MoE agent with about 3B parameters active per token. You can download the weights. Once they run on your hardware, the chat does not need to leave your building.

server rack

Alibaba’s Accio team put Occamy-1.0 on Hugging Face under Apache 2.0: model weights, some training data, and the Dressage tools used to train it. It is built for “co-work,” meaning long jobs that call tools, edit files, run code, and keep going across many steps.

Why the size math matters for hosting

Occamy starts from Qwen3.6-35B-A3B. It has 35 billion total parameters. Only about 3 billion are active for each token, because it is a mixture-of-experts model. That is the hosting point. You get agent-style behavior without paying frontier prices for every small step in a long workflow.

The team’s own card points to SGLang or vLLM with eight-way tensor parallelism and a very long context window. That is real engineering work. It is not a one-click consumer app. For teams that already serve models, it is a known path. For everyone else, it is a project.

What you buy when you host

Self-hosting open weights is a privacy and control choice before it is a benchmark choice.

Your prompts, CRM updates, code, and invoices can stay on your network. There is no round trip to a vendor cloud for each tool call. Cost and latency become yours to manage. For long agent jobs, that can matter more than a few points on a vendor chart.

Once the weights are on your machines, Alibaba does not get a live feed from that run. Security researchers have said there is nothing magic in open weights that must phone home. The trade is different: you own ops, patches, and evals.

What the scores do not settle

Accio reports strong co-work numbers on its own suites, including Claw-Eval, and says Occamy uses fewer tokens and fewer tool calls than its base. Those figures are self-reported. They are not independent audits. Treat them as a sales pitch with a paper attached, then test on your own tasks.

There is also a provenance caveat. Anthropic has alleged Alibaba-linked distillation of Claude’s agent skills. That claim is disputed territory, not a court verdict. Teams that care about training-data lineage should weigh it before they put Occamy on sensitive work.

The wider lesson

Occamy is one release. The pattern is bigger. Open-weight agents make a real fork: rent a frontier API for every step, or host a smaller active model where your data already lives. Sovereignty and privacy show up here as ops, not slogans. If you can run the agent next to the files, you decide who sees the run.

Sources

Source: Alibaba Releases Open-Weights AI Agent Claiming Frontier Co-Work Scores at 3B Active Parameters