Dutch healthcare AI startup Juvoly today announces the open-source release of J1, its new 8-billion-parameter clinical reasoning model. The full model, including evaluation tools and data curation pipelines, is available on Hugging Face.
A small caveat: J1 currently operates under a Llama 3.1 license, though Juvoly intends to switch to a base model with an MIT license in the next release. This first version, J1, is built on the foundation of Llama 3.1-8B-Instruct and trained on 100 billion synthetic tokens, generated from CC-BY-licensed PubMed articles. The complete training process took only 2,688 GPU hours, made possible by Juvoly’s powerful NVIDIA B200 systems.
The result is an ultra-efficient model that delivers faster inference with significantly lower energy consumption. While J1 does not yet surpass the colossal GPT-4o, it outperforms all leading models of similar size—including Qwen3-8B, HuatuoGPT-o1-8B, and the Dutch Delphyr M1—on key medical benchmarks. J1 achieved a score of 79.34% on MedQA and 81% on PubMedQA, a strong indication of J1’s ability to accurately interpret and answer medical questions.
Although the model excels at clinical reasoning, it remains to be seen how useful it can be in real-world applications, according to Thomas Kluiters, founder and CEO of Juvoly. “Benchmarks like MedQA and PubMedQA provide a reasonable indication of how well a model can answer clinical questions, but they don't measure how suitable it is for performing administrative tasks such as drafting discharge letters and consultation reports, or automatically extracting structured data from patient records. Our next step is to develop benchmarks that better reflect these real-world use cases, so that we—and other AI developers—can train models that truly add value to healthcare.”
To realize this ambition, Juvoly is seeking collaboration with hospitals and other major healthcare institutions. Interested organizations are invited to get in touch through www.juvoly.nl
charliewulff•9mo ago
A small caveat: J1 currently operates under a Llama 3.1 license, though Juvoly intends to switch to a base model with an MIT license in the next release. This first version, J1, is built on the foundation of Llama 3.1-8B-Instruct and trained on 100 billion synthetic tokens, generated from CC-BY-licensed PubMed articles. The complete training process took only 2,688 GPU hours, made possible by Juvoly’s powerful NVIDIA B200 systems.
The result is an ultra-efficient model that delivers faster inference with significantly lower energy consumption. While J1 does not yet surpass the colossal GPT-4o, it outperforms all leading models of similar size—including Qwen3-8B, HuatuoGPT-o1-8B, and the Dutch Delphyr M1—on key medical benchmarks. J1 achieved a score of 79.34% on MedQA and 81% on PubMedQA, a strong indication of J1’s ability to accurately interpret and answer medical questions.
Although the model excels at clinical reasoning, it remains to be seen how useful it can be in real-world applications, according to Thomas Kluiters, founder and CEO of Juvoly. “Benchmarks like MedQA and PubMedQA provide a reasonable indication of how well a model can answer clinical questions, but they don't measure how suitable it is for performing administrative tasks such as drafting discharge letters and consultation reports, or automatically extracting structured data from patient records. Our next step is to develop benchmarks that better reflect these real-world use cases, so that we—and other AI developers—can train models that truly add value to healthcare.”
To realize this ambition, Juvoly is seeking collaboration with hospitals and other major healthcare institutions. Interested organizations are invited to get in touch through www.juvoly.nl