Arcee's Trinity Mini Learns to Search Evidence, Not Do Science

Arcee AI, the drug-discovery startup Loka, and the open-source training group Prime Intellect ran 21 controlled post-training experiments on Trinity Mini, Arcee’s open mixture-of-experts model, teaching it to search biomedical databases with tools and to structure gene-function annotations under reinforcement learning. The promoted checkpoint, called Run 120, raised the held-out Drug Tool score from 70.8 percent to 81.2 percent and reached a combined score of 0.863 on the held-out BioReason evaluation. Arcee describes both benchmarks as its own, built specifically for this project.

Trinity Mini packs 26 billion parameters into a mixture-of-experts design that activates only 3 billion per token, spread across 128 experts (eight routed plus one shared) with a 128k-token context window. The new work builds on an earlier Loka project, Trinity Mini DrugProt-Think, in which the same base model learned to sort drug and protein pairs pulled from biomedical abstracts into relationship categories. Here the assignment expanded: pose an open-ended question, gather evidence from multiple tools despite dead ends and conflicting sources, and hand back a result a human scientist could verify.

The team built two reinforcement-learning environments. Drug Tool grades whether the model picks the right resource out of seven retrieval integrations, a set that spans PubMed, GEO, KEGG, UniProt, and STRING. A separate, optional tier of NVIDIA NIM tools extends that toolkit further, letting the model request molecular generation, ligand docking, or protein-folding predictions when the task calls for them. BioReason instead requires the model to infer Gene Ontology identifiers from protein evidence and return exactly one valid JSON object, with grading built from GO term overlap (F1), similarity across the ontology tree, and whether the output format holds up.

Post-training used GRPO, the reinforcement-learning method introduced in DeepSeek’s DeepSeekMath paper, applied through LoRA adapters rather than full fine-tuning. Prime Intellect’s prime-rl framework separated rollout inference from policy optimization across a four-plus-four GPU topology. Active training context ran to 16,000 tokens, with rollouts capped at 7,168 completion tokens and up to 12 tool-use turns during training, eight during held-out evaluation.

Training data came from two sources. Loka assembled the Drug Tool SFT set from its own drug-discovery prompt bank: 800 prompts for training, another 200 held out for evaluation, spanning 17 distinct workflow categories and totaling 5,049 individual tool calls. Those trajectories, covering both the assistant’s reasoning steps and its tool calls, came out of Arcee’s Trinity Large Thinking model running against OpenRouter’s interface. On the other side, the BioReason corpus contributes 8,630 curated records built on the existing BioReason-Pro benchmark.

Before reinforcement learning, the team ran a single prompt-optimization pass called GEPA on each environment, using Trinity Mini to generate rollouts and a Claude Sonnet 5 model to propose revised instructions. That pass alone lifted base-model validation by roughly 84 percent on the BioReason task and by 7.7 percent on Drug Tool, a separate figure from the final post-training numbers above. The two environments responded differently to that gain: one kept the improvement through reinforcement learning, and the other mostly absorbed it during training, which is why Arcee ran the prompt search once rather than repeatedly around a trained checkpoint.

The Run 120 adapter now powers a companion application, AI Scientist, whose stack combines Strands, FastAPI, and React. Above the specialists sits a single orchestrator that hands work to six named agents, two of which run their own sub-specialists, for eight specialist agents in total beneath it. Python execution for those agents is walled off in its own AWS Bedrock AgentCore sandbox, separate from the main application process.

None of this means Trinity Mini conducts scientific research on its own. Arcee’s own write-up is explicit that these numbers describe performance inside the two training environments and their automated checks. It does not claim the model demonstrated anything about laboratory results, clinical outcomes, or the success of an actual drug program. What the model has actually learned is narrower and more mechanical: choosing the right database tool, recovering when a query fails, and returning gene annotations in a format a downstream system can parse without breaking.

For teams weighing whether to fine-tune an open model for a regulated technical workflow instead of building a bespoke one, the reusable part is the process, not the score: define held-out tests and verifiable rewards around one workflow, train a small LoRA adapter instead of a new foundation model, and review the traces before promoting any checkpoint.

Arcee AI published this account of the project, credited to Loka and AWS, on its blog on July 30, 2026.