How Tamarind is integrating NVIDIA NIM microservices and BioNeMo Agent Toolkit to power autonomous molecular design
The models are here. Along with the explosion in general reasoning, AI is becoming more and more present in how new molecules, whether drug candidates, reagents, diagnostics are discovered and designed.
For the last decade, computational biology meant a researcher picking a model, standing up a job, waiting, interpreting the output, and deciding what to run next. Tamarind was founded to resolve as much of the friction in this process as possible, deploying tools and models in our standardized web interface and API, allowing scientists to focus on how they apply models to their problems as opposed to tedious deployment work, scaling compute or building reproducible pipelines.
We believe the next decade of how computational science gets done will look massively different: AI agents that can design and execute plans end-to-end, apply the correct tools in the right place, and iterate.
Tamarind Bio exists to help biopharma R&D organizations become agent ready. Our role is the neutral infrastructure-layer, allowing any scientist, and increasingly any agent, to run state-of-the-art computational biology tools without managing the GPUs underneath.
Agents only work if the tools they call are fast, reliable, and portable. That is exactly where NVIDIA BioNeMo Agent Toolkit fits into what we're building. A general-purpose reasoning model can read a paper or write code, but it can't fold a protein, score a binder, or generate a viable backbone. Agentic biology depends on specialized, domain-specific models that produce real, computable scientific outputs, and those models carry heavy, GPU-bound inference. NVIDIA NIMs are built precisely for this: optimized, containerized microservices for biomolecular models, exposed through standard APIs and portable across environments.
How we’re using BioNeMo
We are excited to integrate NVIDIA optimized NIM microservices into the inference layer that powers the Tamarind platform. Concretely, this means:
Increased inference speed and efficiency for scientific AI inference
Agent-ready tools. Thanks to the Tamarind MCP server and standardized API schema to a wide selection of molecular AI tools, NIM containers can be incorporated into any agent or LLM.
Deployment that meets biopharma where they are: NIM containers fit into Tamarind’s modular architecture, allowing containers to be used in Tamarind or customer-hosted environments, which in-turn, thanks to Tamarind’s agent and API integrations can be used within ELNs, LIMS, or internal model deployment processes.
Efficient inference and agent skills lead to faster discovery
As agents begin to run long-horizon scientific tasks, as with human scientists, they use results from specialized tools to feed their reasoning. Results from these tools, such as biomolecular structure predictors, often have multi-hour feedback cycles.
This is the bottleneck BioNeMo addresses. The inference optimizations baked into BioNeMo microservices make it feasible to meet the heavy inference demands of a platform like Tamarind, which runs tens of thousands of inference calls in the course of agents' scientific reasoning.
The value of that speedup compounds rather than adds. When a human scientist submits a folding job, the latency is a one-time cost paid in parallel with everything else they're doing. When an agent submits one, that latency sits directly on the critical path of its reasoning: it cannot choose its next action until the result returns. Stretch that across a campaign of dozens or hundreds of dependent steps and per-call latency stops adding up and starts multiplying into the wall-clock length of the entire campaign. A 5x faster structure predictor (e.g. AlphaFold2) isn't 5x faster on one prediction; it's 5x across every link in a long chain of decisions.
We are thrilled to be a launch partner for NVIDIA BioNeMo Agent Toolkit, TODO
About Tamarind Bio
Tamarind Bio is a leading provider of 300+ molecular AI tools, providing a web interface, API, and agent to open-source, internal, and third-party proprietary models. In addition to deploying models at massive scale, the platform supports connecting multiple tools together into pipelines, fine-tuning them on internal custom data, and more.
Tamarind is deployed in the majority of the top 20 pharma, supporting tens of thousands of scientists across industry and academia. The platform is applied in a diverse array of tasks including protein design, computational chemistry, and molecular property prediction.