Large language models (LLMs) are increasingly considered as adjuncts for clinical reasoning support, yet most evaluations have focused on cloud-hosted proprietary models. Their open-weight counterparts, which can be deployed locally without transmitting patient data to external servers, remain underexplored—particularly for neglected tropical diseases (NTDs). This exploratory study evaluated the diagnostic performance of seven open-weight LLMs (Falcon 180B, Phi-3 mini 3.8B, LLaMA 2 70B, Meditron 70B, Mixtral 8x7B, DeepSeek-V2 16B, and DeepSeek-R1 70B) for visceral leishmaniasis, using eight clinical vignettes identical in content, prompt, and presentation order to those previously employed to evaluate ChatGPT/GPT-4 by our group. All models were served locally via the Ollama runtime under default inference settings. Top-five and top-one diagnostic accuracy were calculated, and a five-member specialist committee (three infectious-disease physicians and two clinical-diagnosis faculty) classified every generated hypothesis as plausible, non-plausible, or fabricated, after independent assessment and consensus adjudication. Top-five accuracy ranged from 100.0% (DeepSeek-R1 70B; 95% CI: 67.6–100.0) to 12.5% (LLaMA 2 70B and DeepSeek-V2 16B). DeepSeek-R1 70B exceeded the previously published ChatGPT/GPT-4 benchmark of 75.0%; Falcon 180B matched it. Three qualitatively distinct failure modes were identified: fabrication of nosological entities (Phi-3 mini), generic vagueness (LLaMA 2 70B), and non-engagement with the diagnostic task (Meditron 70B). These findings demonstrate that locally-deployed open-weight LLMs can match or exceed proprietary models for VL differential diagnosis, while underscoring the need for expert-validated qualitative assessment beyond binary accuracy metrics. Institution-level evaluation should precede clinical adoption.