A team does the MIT agent test again on four RTX 3090 GPUs

A research team that works with Apex Fusion does the MIT SwarmWorld test again on four RTX 3090 GPUs. MIT put 50 or more AI agents in a shared simulated world. The test took about 26 hours and 12,800 model calls. Nine of 12 paired effects matched MIT in direction. The team used the MIT code and commands with no change.

How it works

The MIT agents had no roles and no recipes. They explored, built technology and wrote programs that continued to operate. The team used Qwen3.8 27B, a model with public weights, compressed to 4 bit precision.

The numbers

Alone, the model reached 6 of 16 research milestones. In a group, it reached all 16. The MIT frontier model reached 14 in a group. The three other effects were in the range of noise.

What is not known

The result comes from one test with one number of agents. The team did not test the larger groups of the MIT study. The team found that KV cache precision changed results more than the difference between a 27B and a 235B model.

Insight: The test also shows a product

Apex Fusion ran the test. It sells Vector, a settlement layer that records which agent did which work on a blockchain. It did a pilot of 32 calls through Vector. It is not known if agent networks must have a blockchain for this record. This reads as a bid to own the record of agent work, with cheap agent groups as the proof. Look for the full test through Vector.