Institute Model
In DevelopmentLucidiModel
LucidiModel is the Institute's own model: an open-weight foundation model, fine-tuned to specialize at multi-agent work. The Institute already runs as a system of agents — orchestration, autonomous research runs, and the collective that produces our papers. LucidiModel is the engine built to drive that system: owned, transparent, and tuned on the Institute's own multi-agent traces rather than rented from a black box.
Why a model of our own
The Institute's research is multi-agent at its core, from million-agent market simulations to the agentic operations that run the lab. A general chatbot is not built for that. LucidiModel is specialized for orchestration: planning, delegation, tool use, and disciplined self-critique. Owning the weights keeps the cost of long autonomous runs low and keeps us honest about exactly how the work is produced.
What it is specialized for
Orchestration & delegation
Decompose a goal, delegate to sub-agent roles (explorer, architect, reviewer, skeptic), and synthesize their results into one coherent answer.
Tool use & structured output
Reliable function calling and schema-constrained output, so the model can act inside real pipelines instead of only producing prose.
Adversarial verification
Falsify before it accepts: source every load-bearing premise, attack its own claims, and refuse conclusions that rest on unverified assertions.
Long-horizon state
Hold and compact context across long runs, so an agent loop can work toward a goal for hours without losing the thread.
How we are building it
We do not pretrain. LucidiModel starts from a leading open-weight base model (Apache-2.0 class, bilingual EN/中文, strong tool use), chosen by a dated, reproducible benchmark rather than reputation. From there we post-train on the Institute's own data: multi-agent trajectories, verifiable-reward signals, and distilled traces from a frontier teacher.
Phase 0 — Base & benchmark
Select the open-weight base, stand up serving, and build LucidiBench: an internal suite of multi-agent tasks scored against verifiable outcomes. Establish a prompt-engineered baseline before spending compute on training.
Phase 1 — Distill & LoRA
Distill traces from a frontier teacher and fine-tune with LoRA on the Institute's own multi-agent runs. A cheap, fast first model whose gains over the baseline are measured, not assumed.
Phase 2 — Verifiable-reward RL
Reinforcement learning where the reward is real: did the simulation reproduce, did the Lean proof check, did the output satisfy its schema. These signals are hard to fake and rare to have — and the Institute generates them every day.
Phase 3 — Deploy & integrate
Quantize for local serving and wire LucidiModel in as the default engine for the Institute's agentic operations. Publish a methods paper documenting what worked and what did not.
LucidiModel is in active development. Naming of the specific base model and benchmark results will be published here as the work is finalized, alongside the methods paper.