Biological AI has two deployment problems that are easy to conflate. First, teams need evaluations that measure whether a model can reason from evidence instead of producing a plausible answer. Second, models that perform well still have to fit the hardware and workflow where biological data is collected.
Capicú has released two open projects that address those problems from opposite ends of the stack: BioManufacturingBench and Cellpose-SAM WQuant.
BioManufacturingBench: reasoning with an answer that can be scored
BioManufacturingBench v1.0.0 is a 2,000-item benchmark for evidence-grounded biomanufacturing reasoning. It contains 1,600 text records and 400 microscopy records across six task families:
- evidence extraction;
- table and unit reasoning;
- process trajectory reasoning;
- microscopy count-range estimation;
- instruction following; and
- abstention when the supplied evidence is incomplete.
Every primary score is produced by a deterministic rule. No primary score uses an LLM judge.
The text portion draws from 500 held-out open-access articles published in 2026. The microscopy portion indexes 3,927 reserved images. Source groups used in the benchmark are excluded from the related training releases, preserving a clear separation between training material and evaluation evidence.
The first public baseline used a stratified 100-item sample across 16 open-weight checkpoints. Assigned-task scores ranged from 43.0% to 96.0%, with especially wide separation on unit-preserving calculations and microscopy. That pilot shows how the benchmark behaves; it is not presented as a definitive ranking from 100 sampled items.
The important result is the separation between capabilities. A system can extract evidence correctly and still fail a mass balance. It can follow a strict output format and still miss the process diagnosis. Reporting one aggregate score without those distinctions hides the failures that matter in technical work.
Cellpose-SAM WQuant: deployment trade-offs made explicit
Cellpose-SAM is a capable generalist model for cellular segmentation, but its original weight footprint can be difficult to place beside a microscope or inside a constrained imaging system.
Cellpose-SAM WQuant publishes five ready-to-load quantized variants:
- W8A16 at 295.8 MiB and 3.93x compression;
- W8A8-QDQ at 295.8 MiB and 3.93x compression;
- mixed-W4-sensitive-W8 at 171.9 MiB and 6.76x compression;
- W4A16-G64 at 168.5 MiB and 6.90x compression; and
- W2A16-G64 at 96.3 MiB and 12.08x compression.
Each model repository includes a machine-readable result and registry record. On the initial 20 paired evaluation samples, W8A16 and mixed-W4-sensitive-W8 met the complete bounded-panel criterion. The more aggressive variants remain available with their limitations visible so teams can evaluate the size-and-fidelity trade-off against their own hardware and data.
Evaluation and deployment belong together
BioManufacturingBench asks whether a model's answer can be supported and scored. Cellpose-SAM WQuant asks what happens when a capable biological model is compressed toward an instrument-sized footprint.
Publishing both makes the development path inspectable: define the task, preserve a governed evaluation boundary, report failures by capability, and expose the trade-offs introduced by deployment constraints.