AI
Alibaba's 3 Billion Parameter Agent Outworks Systems Many Times Its Size
The Accio Team inside Alibaba Group released Occamy 1.0, an open weights AI agent built for co work, the sustained multiple step digital labor that modern enterprises actually buy. updating records, running finance workflows, writing and executing code, editing files, and recovering from errors across dozens or hundreds of tool calls. Only about 3 billion of its 35 billion parameters activate per token, thanks to a mixture of experts design, which keeps the inference footprint small enough for far more accessible hardware.
The team reports headline numbers that punch above the model's weight. On Claw Eval, the paper's main co work benchmark, Occamy 1.0 scores 82.20 on the average metric, ahead of GPT 5.6 Sol at 81.80 and DeepSeek V4 Pro at 81.70, though behind Qwen 3.8 Max at 83.92. Its pass at 3 score of 71.40 also tops GPT 5.6 Sol's 68.90. On Business Arena, a simulation of managing a real business through extended workflows, it reached a final simulated net worth of $79,868, the best result among comparable scale models.
The efficiency story may matter more than the trophies. Compared with the Qwen 3.6 35B A3B base it started from, Occamy 1.0 uses 19.5 percent fewer tokens per task trajectory, makes 15.2 percent fewer tool calls, and cuts wall time by 46.4 percent. Its task timeout rate fell from 9.88 percent to 2.18 percent. In agentic work, where cost and latency accumulate across every step of an episode, fewer dead end calls translate directly into lower bills.
Everything needed to reproduce the work ships openly. the weights, a subset of training data, and the Dressage training framework, released under open terms. The technical paper documents a three stage post training recipe, supervised fine tuning across nearly 15,000 trajectories, and a model soup merge of two expert checkpoints into one release.
Two caveats deserve daylight. Every benchmark figure comes from the Accio Team's own paper and model card, and independent replication has yet to land, so treat the table as the team's best case until outside labs confirm it. Anthropic also alleged in June that Alibaba linked operators ran 28.8 million Claude exchanges through about 25,000 fraudulent accounts to distill agentic capabilities; Alibaba has stayed silent on the claim, which remains attributed to Anthropic, with no independent resolution yet.
The honest read is that a compact open agent genuinely competing with frontier systems on the work that costs money is a gift to every team running AI at scale, and the open training stack means the ideas travel even to builders who skip the weights. Run your own task specific evaluation before trusting any benchmark table, including the one in this story.
Quick answers
What is this story about?
The Accio Team inside Alibaba Group released Occamy 1.0, an open weights AI agent built for co work, the sustained multiple step digital labor that modern enterprises actually buy. updating records, running finance workflows, writing and executing code, editing files, and recovering from errors across dozens or hundreds of tool calls. Only about 3 billion of its 35 billion parameters activate per token, thanks to a mixture of experts design, which keeps the inference footprint small enough for far more accessible hardware.
Why does this story matter?
The honest read is that a compact open agent genuinely competing with frontier systems on the work that costs money is a gift to every team running AI at scale, and the open training stack means the ideas travel even to builders who skip the weights. Run your own task specific evaluation before trusting any benchmark table, including the one in this story.
Sources
- Tech Times: Alibaba releases open-weights AI agent Occamy 1.0
- Hugging Face: Accio-Lab Occamy 1.0 model card
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.