[{"data":1,"prerenderedAt":101},["ShallowReactive",2],{"profile":3,"writing-small-models-big-deployments":14},{"name":4,"role":5,"tagline":6,"location":7,"contact":8},"Sankar Vema","AI Builder & Architect of Agentic Systems","I help enterprise leaders turn AI ambition into capability that actually ships.","India · open to global advisory engagements",{"email":9,"linkedin":10,"github":11,"blog":12,"twitter":13},"sankar.vema@gmail.com","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Fsankarvema\u002F","http:\u002F\u002Fsankarvema.github.io\u002F","http:\u002F\u002Fsankarvema.blogspot.com\u002F","https:\u002F\u002Ftwitter.com\u002Fsansvema",{"id":15,"title":16,"body":17,"date":86,"description":23,"draft":87,"excerpt":88,"extension":89,"meta":90,"navigation":91,"path":92,"seo":93,"slug":94,"status":88,"stem":95,"tags":96,"__hash__":100},"writing\u002Fwriting\u002Fsmall-models-big-deployments.md","Small Models, Big Deployments",{"type":18,"value":19,"toc":78},"minimark",[20,24,29,32,35,39,42,45,49,72,75],[21,22,23],"p",{},"There is a reflex in most AI projects: reach for the biggest, most capable model\navailable, and route everything through it. It feels safe. It is often the wrong\ncall. A large share of real enterprise work does not need frontier intelligence.\nIt needs a competent model that is small, fast, cheap, and running somewhere you\ncontrol.",[25,26,28],"h2",{"id":27},"the-case-nobody-makes-in-the-demo","The case nobody makes in the demo",[21,30,31],{},"Demos are built to impress, so they use the largest model. Production is built to\nsurvive, and production has different priorities: latency, cost per call, data\nresidency, and reliability under load. On every one of those, a well-chosen small\nmodel, sometimes running on your own hardware or even on the device, quietly wins.",[21,33,34],{},"Classifying a document, extracting fields, routing a ticket, drafting a\ntemplated reply, checking a form: these are high-volume, bounded tasks. A small\nmodel fine-tuned on your data does them faster and cheaper than a giant\ngeneral-purpose model, and it does not send your sensitive data to someone\nelse's cloud to do it.",[25,36,38],{"id":37},"why-this-matters-more-now","Why this matters more now",[21,40,41],{},"Two forces make small models a serious strategy, not a compromise. Small models\nhave become genuinely capable, closing much of the gap on narrow tasks. And the\ncost and privacy pressures of running everything through a frontier API have\nbecome real line items on the P&L and real questions from your risk committee.",[21,43,44],{},"The mature architecture is not \"one big model for everything.\" It is a portfolio:\na small model for the routine high-volume work, a large model reserved for the\ngenuinely hard reasoning, and a router that sends each request to the cheapest\nmodel that can do the job.",[25,46,48],{"id":47},"what-to-do-now","What to do now",[50,51,52,60,66],"ol",{},[53,54,55,59],"li",{},[56,57,58],"strong",{},"Audit your traffic."," Most requests are routine. Measure what share truly\nneeds frontier reasoning. It is usually smaller than the team assumes.",[53,61,62,65],{},[56,63,64],{},"Default small, escalate to large."," Make the big model the exception you\nreach for, not the default you pay for on every call.",[53,67,68,71],{},[56,69,70],{},"Treat privacy as architecture."," For sensitive data, a model you run is not\na nice-to-have. It is the difference between a yes and a no from compliance.",[21,73,74],{},"Bigger is not a strategy. It is a default, and defaults are where money and trust\nquietly leak.",[21,76,77],{},"So before your next model decision: how much of your workload is paying frontier\nprices for routine work?",{"title":79,"searchDepth":80,"depth":80,"links":81},"",3,[82,84,85],{"id":27,"depth":83,"text":28},2,{"id":37,"depth":83,"text":38},{"id":47,"depth":83,"text":48},"2026-08-01",false,null,"md",{},true,"\u002Fwriting\u002Fsmall-models-big-deployments",{"title":16,"description":23},"small-models-big-deployments","writing\u002Fsmall-models-big-deployments",[97,98,99],"ai","architecture","cost","7pN_-FqeRav1zu11LQj0Ya4WnteHNdZfpd3zR-Dquj0",1790601541709]