On Oct. 23, 2026, @notanilp, who wrote that they started Mubit in Aberdeen, announced three offerings: Mubit, a platform pitched as turning agents' work into models customers own; Banyan Sapling, an open model; and latte/bench, an agent benchmark.
On its site, Mubit describes a service that builds dedicated models for agents' recurring tasks while retaining existing frontier models. The company says it switches to a dedicated model only when that model does better. Its FAQ also describes using agent runs to propose changes to memory, prompts and skills. The site invites visitors to join a waitlist and advertises a launch week of Oct. 26-30, 2026.
The Oct. 23 post offered $100 in credits to people who commented “SAPLING” within 48 hours and advertised a free first month for enterprises.
What Sapling and latte/bench test
MubitLabs' model card calls Banyan Sapling a research preview for checking whether a claim is supported by supplied policy and evidence. It describes outputs for conflicts, missing information and inconsistent context, while leaving retrieval and source authentication to the surrounding workflow. The card says the model weights are distributed under Apache 2.0.
The card reports 94.22% raw accuracy across 1,731 evaluation cases, but calibrated answers on only 779, or 45.00%. In the 731-case ContractNLI partition, it reports calibrated answers for six cases, or 0.82%. MubitLabs cautions that these results do not establish real-world accuracy or independently estimate performance on unseen organizations.
Mubit's latte/bench description has models plan simulated latte pours, receive feedback and try again, comparing attempts with and without a retained note. It says the study tests planning and adaptation through client-side notes, not model-weight learning, held-out transfer or real-world pouring skill.
Comments