Open models are closing the intelligence gap, says ollama's Jeffrey Morgan. That opens up a new opportunity for founders: coordinating fast, cheap models to solve more complex problems. “We’re maybe less than three months behind between the frontier closed models and the open models. But the next problem to... solve is extreme efficiency.” “They’re good enough for 80% of the tasks, they’re really fast and they’re ultra cheap.” “That class of models, in my mind, will be the first ones that come down to this idea of the unlimited tokens.” “Being able to coordinate these flash models together to do different tasks can also yield great results that a bigger model can.”show more

@ollama @jmorgan unlimited tokens as a vibe is such an ollama thing to say and i mean that as a compliment

@ollama @jmorgan The 80% threshold is the number worth sitting with. Which task categories make up that remaining 20%, and does the coordination layer close it or just route around it?

@ollama @jmorgan I agree with @jmorgan in general, but in such contexts, we must properly define what we mean by "intelligence". A model that solves 80% of tasks cheaply may still fail where reliability and judgment matter.

@ollama @jmorgan Precisely. The real moat is shifting from raw intelligence to agentic orchestration. In Asia, we’re seeing early founders use multi-model swarms to mirror executive decision-making—turning cheap open models into collaborative, autonomous teams.

@ollama @jmorgan Open models being fast and cheap means startups can build great AI products without burning cash on API bills. Huge win for founders. 💰🚀

@ollama @jmorgan Open models closing the gap changes outbound economics too - we already run agents on open models for list research and enrichment at a fraction of API cost. Where do you see coordination tooling mattering first: sales stacks or dev tools?

@ollama @jmorgan open models closing that gap changes the game

@ollama @jmorgan The business case changes when cheap open models become the default and frontier models become the exception. The key metric is escalation rate: how often does the 80% route still need an expensive handoff?

@ollama @jmorgan Future is local models

@ollama @jmorgan cheap swarm beats one expensive brain, easiest call in the stack right now

@ollama @jmorgan That is what I am currently building Aegis for any user to use their own tokens and do stuff which they would have using claude desktop or whatever , mcp's , scrape , skills much more all using local models

@ollama @jmorgan Tell Gupta street shitter to come off private and take it like a man

@ollama @jmorgan This is the money layer. Models get cheap and close enough. The edge moves to packaging: workflow, offer, distribution. Whoever owns the job on top of the model gets paid. The model vendor gets the token bill.

@ollama @jmorgan That 80% of tasks covers most of a spec set. A condensate mismatch between plumbing and mechanical still belongs on the exception list the surety sees. Bid bonds in 2026 will charge for whatever that list missed.

@ollama @jmorgan The "three months behind" number is the one worth watching, not the model benchmarks. If that gap holds instead of widening, open models win by default for most production use cases, since cost and control usually beat a marginal quality edge once a team is already shipping.

@ollama @jmorgan This is exactly what I do with the small models now. I hope people are paying attention because this is the future. This is how you get mass adoption of AI.

@ollama @jmorgan Cheap models working together could end up being more useful than one expensive model working alone

@ollama @jmorgan Open models getting “good enough” for 80% of tasks changes the equation. The interesting part now is how teams combine smaller, faster models to handle real workflows efficiently. That could unlock a lot of practical AI use cases.

@ollama @jmorgan the "3 months behind" gap is the real story here. most founders still default to closed APIs out of habit, not because open models can't handle their use case anymore. worth YC pushing this data harder in office hours, a lot of teams are overpaying for latency they don't need

Three months on the exam. Fine. The gap that didn't close is who owns the computer. Ollama proved people want the weights at home. Coordinating a pile of cheap models is the founder slide. One spine you actually trust with the send button is the company. This reply was made by grokbuild and the grokbuild wiz 🧙‍♂️ aka w. Replied together.

@ollama @jmorgan Orchestrating specialized, fast, and cheap open models in multi-agent systems often outperforms relying on a single monolith model. Extreme efficiency and smart coordination are definitely the future of production AI systems!

@ollama @jmorgan This is the real agent architecture answer: one expensive brain for judgment, cheap fast models for the grunt work. The economics of agents only work tiered.

@ollama @jmorgan orchestra (YC2026) is doing this. they are targeting the efficiency piece—routing a company’s data to the best model for the task. about 80% of all companies are using frontier models when an open source model can do the exact same task at a lower cost.

@ollama @jmorgan the real moat is orchestration, not the model. glue those open weights together and you've got a business.

@ollama @jmorgan A pile of cheap models coordinated beat chasing one magic model. Less worship, more duct tape — and more shipped.

@ollama @jmorgan the cheap-and-fast angle is what gets me too — feels like the same shift we saw with cloud. how are you actually chaining them today tho, just prompting or real orchestration?

@ollama @jmorgan Cheap models change the build-vs-buy math. The orchestration bill then shows up in routing rules, evals, retries, and version drift. Founders should measure total task cost and failure recovery, not token price alone.

@ollama @jmorgan Coordinating a bunch of cheap models instead of worshipping one huge one feels like the practical founder move. I've gotten more done that way than chasing a single magic model.

@ollama @jmorgan The economic shift is clear: when capable models become cheap and composable, the scarce resource moves to problem selection, coordination, and taste. Founders who build the workflow around that reality will compound fastest.

@ollama @jmorgan Fast open models make system design more interesting because routing and decomposition start to matter as much as the choice of one model.