Build 02Placeholder
30B-class coding assistant on CPU
One developer, a 30B-class code model, no GPU required to start.
A CPU machine with enough RAM to hold a 30B-class quant and a prompt long enough to be useful in a repo. Speed is unmeasured on purpose.
- Price
- Placeholder
- Price pulled on
- Placeholder
- Source
- Placeholdername the sold listing or retailer page on the day it is pulled
- Idle
- Placeholder
- Load
- Placeholder
- Model
- Placeholder
- Quant
- Placeholder
- Context
- Placeholder
- Gen tok/s
- Placeholder
- Prompt tok/s
- Placeholder
Wall watts under the coding-model run, plus idle. Spec-sheet TDP stays off this row.
Prompt speed matters here as much as generation. A long file is a prompt-processing problem.
Don’t buy this if
- You already have a 24GB GPU that holds the quant you want. That build is the GPU page, after you measure it.
- You need a pace you have not seen measured, and you are unwilling to wait on a CPU.
- The listing says workstation RAM, and the photo shows one DIMM or a board that cannot run the channels it claims.
Parts
| Part | Item | Price (USD) | Source | Notes |
|---|---|---|---|---|
| Platform | Placeholdera desktop or workstation with real memory bandwidth | Placeholder | Placeholder | More cores are not automatically more tokens per second. Bandwidth is the thing to suspect. |
| Memory | Placeholdercapacity for a 30B-class quant, cache, and OS headroom | Placeholder | Placeholder | Context length spends RAM after the weights are loaded. Record both. |
| Storage | PlaceholderSSD for several mid-size model files | Placeholder | Placeholder | You will try more than one quant. Leave room. |
| Cooling and power | Placeholdera PSU and cooler that match a sustained CPU load | Placeholder | Placeholder | Coding sessions are long. A box that throttles after a minute is a different machine. |
No affiliate links on this page yet. If one is added, the source cell will say so. See the affiliate disclosure.
Gotchas
- The quant is part of the result. A smaller quant fits easier and answers differently. Write the quant in the benchmark row.
- A model that loads at a short context can fail at the context you wanted for a repository. Test the length you will use.
- CPU inference is often limited by memory bandwidth. A high core count on a narrow memory bus is a common used-market story.
A coding assistant is a harder job than short chat. The prompt is a file, or several files, and the wait before the first token is the part you feel. That is why this page tracks prompt tokens per second next to generation tokens per second. One number would hide the slow half.
Thirty billion parameters is a class, not a model. Two quants of two different coders can disagree on quality and on whether they fit. This page will not pick a winner in advance. The first entry in the benchmark table that names a real coder, quant, and context replaces the placeholders above.
CPU-only is the point of the objective. It keeps the price and the power about the platform you can actually buy used, without a scarce GPU in the total. If a later measurement says the pace is wrong for daily use, the upgrade path is the used 24GB GPU, added to a host that already has the RAM. It is not a second guess published as a speed.
Prices stay empty until a listing is pulled and dated. A “around this much” figure would age into a lie, and the page would have no date to convict it.
Models that point here
A judgment from the model card and the size, not a speed ranking. Memory and the measured-or-not speeds are on each model page.
- Qwen3-14B · 14.8B
- Phi-4 · 14B
- DeepSeek-R1-Distill-Qwen-14B · 14B