Your quota is exhausted. Local models, hybrid use.
Token prices fall, consumption rises, quotas die faster. I moved toward local models on a high-RAM machine. Hybrid: keep cloud for the hard parts. You do not need a god model for a CRUD.
Your quota is exhausted when cloud token use outruns the plan, even as average token price falls. Consumption only goes up. As prices drop, more use cases become feasible. They get considered. Consumption rises.
Month after month my workflows get more complex and more autonomous. Quotas run out faster and faster. I do not think I am the only one. I feel dependent. Going back on this way of working is no longer possible.
Canonical URL: https://www.zheat.xyz/en/insights/quota-exhausted-local-models/
Table of contents
- Why do quotas die faster?
- What is left when the quota is gone?
- GPU vs MacBook / DGX Spark
- Which local model did I land on?
- What is hybrid use?
- FAQ
- Related
Why do quotas die faster?
Average token price goes down and consumption only goes up. As prices drop, use cases become feasible. They get considered. Consumption rises.
Even providers struggle with demand. Ollama, for example, closed sign-ups to Max plans.
What is left when the quota is gone?
One remaining path: your own machines.
The mistake, I think, is always wanting the best model for our tasks. Today you can work with local models and get results that are good enough.
Two things matter: how much RAM you have, and bandwidth.
GPU vs MacBook / DGX Spark
Two kinds of machines:
| Machine | RAM vs bandwidth | Who it is for in this note |
|---|---|---|
| GPUs | More expensive for the RAM size, very fast bandwidth | The Rolls-Royce |
| MacBook or DGX Spark | More RAM, slower bandwidth | What I picked, because I am not that rich (or rich at all) |
Which local model did I land on?
On Spark-class machines, MoE models are especially interesting. I landed on Qwen 3.6 35B A3B NVFP4. Best compromise I found for speed versus intelligence.
| Model | What happened |
|---|---|
| Qwen 3.6 35B A3B NVFP4 | Best compromise I found for speed versus intelligence |
| GLM-4.7-Flash-AWQ-4bit | Stability issues |
| Qwen3.8-27B-NVFP4 | Too slow for daily use. I might keep it for occasional planning. |
Benchmarks for these models are on Spark-Arena. If you have a recent MacBook and want to try, you should get similar performance.
What is hybrid use?
Hybrid use means you keep cloud models for what is actually hard, and you run the rest local. You do not need Mythos, Zeus, or God to do a CRUD and send it to the front.
The goal for the next months: that split. Other upside: token economy no longer applies to you, and the models do not stop when the quota is gone.
If you think local models cannot code, you are wrong.
This pairs with parallel agents are a false good idea: local capacity is lower, and that can be the point.
FAQ
Why do AI quotas run out faster even when token prices fall? Average token price goes down and consumption only goes up. Cheaper tokens make more use cases feasible. They get considered. Consumption rises. Workflows get more complex and more autonomous. Quotas die faster.
Can local models replace cloud when the quota is exhausted? They can cover a lot of daily work with results that are good enough. Keep cloud for what is actually hard. You do not need a god model for a CRUD.
GPU or MacBook / DGX Spark for local models? GPUs are faster on bandwidth and more expensive for the RAM size. MacBook or Spark-class machines have more RAM and slower bandwidth. I went with the second because I am not that rich.
Which local model did you pick? Qwen 3.6 35B A3B NVFP4, as the best speed-versus-intelligence compromise on Spark-class machines. GLM-4.7-Flash-AWQ-4bit had stability issues. Qwen3.8-27B-NVFP4 was too slow for daily use.
Where are the benchmarks? On Spark-Arena. A recent MacBook should get similar performance.