A local coding model for each memory size
Memory decides which model fits. We have measured 3 coding classes: Apple silicon 64–128 GB, one NVIDIA L4 24 GB, and Linux CPU 8–32 GB. Every other size gets the arithmetic and the gap, because a fit is not a measurement.
What has to fit
A run needs room for the weights, the context cache, the host, and the harness. On Apple silicon and a CPU-only Linux box that room is system RAM. On a discrete GPU the weights and context sit in VRAM, and the host still needs system RAM.
The catalog lists a minimum RAM for each setup, and a minimum VRAM where one applies. A fit tells you the run can start. It does not tell you how fast it will answer, and a model that fits on a small CPU can still be slow.
- Weights: the model file the setup pins.
- Context cache: grows with the context window the manifest selects.
- Overhead: the host process and the harness, which is why a fit with no headroom is not a fit.
The table, by memory class
A row without a measured record says so. It does not borrow a number from a class that was measured.
| What fits by arithmetic | What we measured | The setup | The gap |
|---|---|---|---|
| 8 GB: the floor for the Linux CPU class. The weights take 1.1 GB and the measured peak was 5.82 GB, which leaves about 2 GB for the OS. | A 7 GB Linux x64 container, below the floor: 157.3 s cold, 119.2 s warm. | Linux CPU page | No record at exactly 8 GB. |
| 16 GB: about 10 GB above the measured peak, and the model stays the same 1.7B. | Not measured at 16 GB. | Linux CPU class page | Lists 16 GB; the record is from a 7 GB box. |
| 24 GB: a single 24 GB card holds both measured L4 models. Their VRAM peaks run from 9.21 GB to 19.49 GB by host. | 3 configurations on one L4 with 15 GB of system RAM. | NVIDIA L4 page | No other 24 GB card, and AWS was not tested. |
| 32 GB: a CPU run has no VRAM pool, so the model stays the 1.7B and extra RAM buys headroom, not a larger model. | Not measured at 32 GB. | Linux CPU class page | No 32 GB record, and no 32 GB Apple record either. |
| 64 GB: 9 GB of weights plus a 32k context fit in unified memory with room to spare. | Tested on 128 GB, not 64 GB. | Apple silicon page | The 64 GB half of the class is unmeasured. |
| 128 GB: the most headroom of any class in the catalog. | 123.6 s cold, 125.4 s warm, 9.69 GB peak on an M4 Max with 128 GB. | Apple silicon page | One machine and one model. |
| 192+ GB: no catalog class. More memory buys headroom; a larger model is a separate run. | Not measured. | The catalog | No page without a passing record. |
The measured numbers come from the qualification records linked on each machine page. The arithmetic rows do not.
Why there is no ranking here
There is no speed ranking here, and no order between the classes. The only claim we can support is that a class has a passing record, and the table marks the classes that do not.
That is why 64 GB stays a gap. The 64 GB Apple machine shares a class with the 128 GB machine we tested, and the record is from the 128 GB one. Scaling its time down would invent a number.
Why the rows name configurations, not models
A model name alone does not answer the question. The same 8B model at a different quantization, context window, or host has a different memory peak and a different run. The catalog pins all of them, so a measured row is a configuration.
The same rule applies to a setup you build yourself. Change the quantization and you have a new memory peak; the record's number no longer applies.
What to do with a gap
Start from the class that matches the machine, read the measured row, and check the listed constraints. If the page says the class was tested on a different size, treat the other size as untested.
A larger model, or a size the catalog never ran, needs its own qualification on real hardware.
Install
Install the CLI, then pick the setup for your machine class.
curl -fsSL https://oloop.run/install.sh | shQuestions
Is 8 GB enough for a local coding model?
The catalog lists a 1.7B Q4 setup at 8 GB, and the measured row came from a 7 GB container with a 5.82 GB peak. A larger model needs a different run, and we have not measured one at 8 GB.
Does more VRAM always mean a faster run?
No. VRAM decides whether the model and its context fit on the card. Speed also depends on the host, the model, the quantization, and the CPU around the GPU.
Why is there no 64 GB VRAM row?
No setup in the catalog has a passing record at that VRAM size. A row without a record would be a guess.
What does 'listed for' mean?
It is the catalog's compatible range, not the tested machine. Each machine page states the tested hardware separately.