{"slug": "loka-and-arcee-detail-an-open-biomedical-model-built-for-aws-deployment", "title": "Loka and Arcee detail an open biomedical model built for AWS deployment", "summary": "Loka, Arcee AI, Amazon Web Services and Prime Intellect are collaborating on open biomedical models for scientific research, with the first result being Trinity-Mini-DrugProt-Think, a LoRA adapter that classifies 13 categories of drug-protein relationships from PubMed abstracts. The adapter, built on Arcee's 26-billion-parameter Trinity Mini model, uses reinforcement learning via Prime Intellect's infrastructure and is deployable on AWS SageMaker, though its reported accuracy rewards are experimental and not independently validated.", "body_md": "[Bojan Jakimovski (@Shekswess)](https://x.com/shekswess), Loka's machine learning and applied research lead, [said in a July 30th thread on X](https://x.com/shekswess/status/2082870774243758590?s=46) that Loka is collaborating with [Arcee AI](https://www.arcee.ai/about), [Amazon Web Services](https://aws.amazon.com/) and [Prime Intellect](https://www.primeintellect.ai/) on open models for scientific research.\n\nThe work at the center of that announcement is [Trinity-Mini-DrugProt-Think](https://huggingface.co/lokahq/Trinity-Mini-DrugProt-Think), a biomedical relation-extraction adapter that Jakimovski and [Petar Kalinovski (@theVladChad)](https://x.com/theVladChad) first documented in February 2026. The timing matters: the July 30th post presents the relationship as a new collaboration, while the underlying model, experiments and AWS deployment guide have been public for roughly five months.\n\nJakimovski, who lists an MSc in dedicated computer systems from Ss. Cyril and Methodius University, also teaches at BrainsterNext and serves as an AWS Ambassador. His research focuses on post-training, small language models and distributed computing. That background shaped the project: instead of training a biomedical foundation model from scratch, Jakimovski and Kalinovski adapted an existing sparse mixture-of-experts model and published the training artifacts needed to reproduce the work.\n\n### What the model does\n\nTrinity-Mini-DrugProt-Think is a LoRA adapter for Arcee's Trinity Mini, a model with 26 billion total parameters and 3 billion activated for each token. The adapter classifies 13 categories of drug-protein relationships found in PubMed abstracts, including inhibitors, activators, agonists, antagonists and substrates. It generates a structured reasoning trace before returning a classification.\n\nThe researchers used Group Relative Policy Optimization, or GRPO, a reinforcement-learning method that rewards outputs that can be checked against a defined answer. Prime Intellect supplied the hosted training infrastructure through its [prime-rl framework](https://github.com/PrimeIntellect-ai/prime-rl), while the training environment used Maziyar Panahi's OpenMed packaging of the DrugProt dataset.\n\nThe [public repository](https://github.com/LokaHQ/Trinity-Mini-DrugProt-Think) contains experiment configurations, exported metrics and scripts for deploying the base model and adapter to an AWS SageMaker real-time endpoint. AWS's role is concentrated on deployment and hosting. The training runs were conducted through Prime Intellect's infrastructure.\n\nThat division of labor makes the project a compact example of the stack emerging around specialized open models: Arcee supplies the base weights, Prime Intellect handles reinforcement-learning infrastructure, Loka builds the domain adapter, and AWS provides a controlled production environment.\n\n### The experiments and their limits\n\nJakimovski and Kalinovski ran 12 experiments across LoRA scaling, learning rates, batch sizes, token budgets, temperatures and rollout counts. Their [technical report](https://lokahq.github.io/Trinity-Mini-DrugProt-Think/) identifies the LoRA alpha setting, which controls the size of adapter updates, as the most consequential variable in the study.\n\nAn alpha of 64 reached an accuracy reward of about 0.75 after 100 steps in one sweep, compared with about 0.41 at alpha 16. A more aggressive alpha of 128 produced stronger training results before its held-out evaluation score deteriorated, prompting the researchers to stop that run at step 70. A learning rate of 1e-5 also collapsed during training, while a 5e-6 run reached roughly 0.83 at step 82 but later showed instability. The researchers selected 3e-6 as the safer default.\n\nThose figures are experimental rewards rather than an independently validated biomedical benchmark. The model card says Trinity-Mini-DrugProt-Think was evaluated on held-out data drawn from the training split and has not been tested against the official DrugProt test set. The report also says the adapter may fail to generalize beyond the specific 13-class classification task.\n\nThe licensing requires similar precision. Loka released the adapter under Apache 2.0, while the underlying [Trinity Mini model](https://huggingface.co/arcee-ai/Trinity-Mini) lists Arcee's separate OpenMDW-1.1 license. Developers deploying the combined system must account for both layers.\n\n### From experiment to deployment\n\nLoka's [February 23rd SageMaker guide](https://medium.com/loka-engineering/deploying-trinity-mini-drugprot-think-on-amazon-sagemaker-ai-9e1c1c430ce9) turns the experiment into a deployable service. The scripts load the immutable Trinity Mini base model, attach the smaller adapter and expose the result through a real-time endpoint. Operators can replace or roll back the adapter without merging a new copy of the full model.\n\nThat architecture targets healthcare and life-sciences teams that need inference to remain inside their own AWS accounts. It gives those teams control over containers, access policies, logs and encryption, while leaving them responsible for GPU sizing, endpoint costs and cold-start delays.\n\nThe renewed attention also follows Arcee's July 22nd announcement of [Genesis-Science-1](https://www.arcee.ai/blog/genesis-science-1), a separate trillion-parameter-class model being developed with the U.S. Department of Energy and its national laboratories. Arcee co-founder and CEO Mark McQuade has shifted the model lab from enterprise post-training tools toward its own open-weight foundation models. The smaller DrugProt project gives Arcee a working example of how outside researchers can specialize those weights for a narrow scientific task.\n\nFor Jakimovski, the immediate output is narrower: an inspectable adapter, its experiment history and a deployment path. Its scientific value will depend on evaluation beyond the training distribution. Its engineering value is already visible in the repository, where other researchers can rerun the ablations, replace the biomedical dataset or carry the same training pattern into another domain.", "url": "https://wpnews.pro/news/loka-and-arcee-detail-an-open-biomedical-model-built-for-aws-deployment", "canonical_source": "https://runtimewire.com/article/loka-arcee-open-biomedical-model-aws-deployment", "published_at": "2026-07-30 21:04:30+00:00", "updated_at": "2026-07-30 21:25:40.641883+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "ai-research", "ai-infrastructure"], "entities": ["Loka", "Arcee AI", "Amazon Web Services", "Prime Intellect", "Trinity-Mini-DrugProt-Think", "Bojan Jakimovski", "Petar Kalinovski", "AWS SageMaker"], "alternates": {"html": "https://wpnews.pro/news/loka-and-arcee-detail-an-open-biomedical-model-built-for-aws-deployment", "markdown": "https://wpnews.pro/news/loka-and-arcee-detail-an-open-biomedical-model-built-for-aws-deployment.md", "text": "https://wpnews.pro/news/loka-and-arcee-detail-an-open-biomedical-model-built-for-aws-deployment.txt", "jsonld": "https://wpnews.pro/news/loka-and-arcee-detail-an-open-biomedical-model-built-for-aws-deployment.jsonld"}}