In early January 2024, the leading closed model beat the best open-weight model by 8.04% on the Chatbot Arena leaderboard. By February 2025, the gap was 1.70%, per the Stanford AI Index (2025). Thirteen months to go from a visible quality tier to a rounding error.
I run open-weight models on my own hardware for research, content operations and automation across my agency. Not as a science project. As default infrastructure, with cloud models reserved for the work that genuinely needs them. After two decades of buying media and two years of building AI automation, I think local-first is the correct default for marketing operations, and the argument has three legs: economics, client data, and control.
The economics moved faster than the debate
The cost curve is the least appreciated chart in AI. Stanford HAI (2025) tracked what it costs to query a model performing at GPT-3.5 level on MMLU.
From $20 per million tokens in November 2022 to $0.07 per million tokens by October 2024: a more than 280-fold collapse in roughly two years (Stanford HAI, 2025).
The same report tracks how small capable models have become. In 2022, the smallest model scoring above 60% on MMLU was PaLM, at 540 billion parameters. By 2024, Phi-3-mini crossed the same threshold with 3.8 billion, a 142-fold reduction, per the Stanford AI Index (2025).
Those two curves together are why local-first stopped being a hobbyist position. Models competent enough for most marketing operations work now fit on a single workstation GPU, and once the hardware sits on your desk, the marginal cost of a token is electricity.
To be fair to the cloud: API prices collapsed too, so raw price alone does not settle the argument. Volume does. Classifying search queries, extracting attributes from product feeds, enriching lead records, summarizing competitor pages, drafting variants for a content pipeline: high-volume, mid-difficulty work that runs all day. On a metered API, someone eventually starts rationing those calls because someone else is watching the invoice. On a machine you own, the correct number of calls is as many as the workflow needs.
Client data is the stronger argument
Marketing operations run on data clients would never want forwarded: customer lists, CRM exports, margin structures, ad account performance. Every cloud LLM call is a transfer to a third-party processor, with whatever contractual and regulatory baggage that carries in each market you operate in.
Practitioners already know the tension. In Cisco’s 2025 Data Privacy Benchmark, a survey of 2,600 privacy and security professionals across 12 countries, 64% worried about inadvertently sharing sensitive information with generative AI tools, and nearly half admitted to entering personal employee or non-public data into them anyway. The concern and the behavior contradict each other, which tells you the tooling, not the awareness, is what is missing.
The same pattern shows up in adoption data: companies cited cost, data privacy and security risks as the top obstacles to their AI initiatives in an S&P Global Market Intelligence survey of more than 1,000 organizations, reported by CIO Dive (2025).
Local models dissolve most of that conversation before it starts. When a client’s CRM export is processed on a machine I own, there is no third-party processor, no retention policy to parse, no cross-border transfer to paper over. On one account in a regulated health niche, sending intake-related data to any cloud model was a non-starter regardless of vendor assurances. A local pipeline did the classification work without a byte leaving the building. That project either happens quietly on local hardware or it does not happen.
Even for cloud-bound work, local models earn their keep as the screening layer. Everything that leaves my machines passes a classification step first, deciding what must be redacted or blocked, and that classifier is local by definition. You cannot outsource the decision about what is too sensitive to outsource.
The third leg is control, and it is the one you only appreciate after being burned. A hosted model can be deprecated, re-priced or quietly swapped for a cheaper variant mid-campaign, and the behavior of every prompt you tuned shifts with it. A local model is pinned: the weights that ran yesterday’s pipeline run tomorrow’s, until I decide otherwise. When a workflow feeds bidding decisions or client reporting, that determinism is not a nicety. It is the difference between a system you operate and a service you hope keeps behaving.
Where local loses, honestly
Frontier reasoning still lives in the cloud. The hardest analysis, long agentic coding sessions and the best writing quality come from closed models, and I pay for them without complaint. Anyone telling you a workstation replaces the frontier entirely is selling GPUs.
Latency cuts both ways too. A local model answers with no network round trip, no rate limit and no queue, which matters when a pipeline makes thousands of small calls. But on a single complex task, a large remote model can still finish faster than a small local one grinding through it.
So the honest architecture is a routing decision, not a religion. Every task gets the cheapest tier that clears the quality bar: local for volume work and anything touching client data, cloud for the calls that genuinely earn their transfer. In my practice, most marketing ops calls never earn it.
Two years ago, running your own models needed an ideological defense. Now it needs a GPU and a weekend. Start local, escalate deliberately, and let the cloud prove it deserves your client’s data instead of assuming it does.