Alternatives to mini-AGI

Bespoke Labs' open-weight 9B model and training recipe that turn Jev-style decisions into single-token choices with probabilities, runnable locally.

Liquid AI's open-weight 2.6B dense model trained for on-device agentic workloads, with 128K context and native tool calling.

Meituan's open-weight MoE model: 1.6T total / ~48B active params, 1M context, MIT. Strong on coding and agentic tasks (2026-08-25).