Is 128GB enough for a local LLM in 2026?

No 128GB machine hands you 128GB. After the operating system takes its share, the model gets 120GB. Every 128GB box on the Grid — Apple and NVIDIA alike — lands on the same usable figure, so the verdicts below hold for the whole class.

At Q4_K_M, one billion parameters occupies 0.604GB. Subtract the runtime overhead and a 128GB machine stores a model of about 196B parameters before any context is counted. That is the whole answer in one number, and everything below is what it means.

“Tight” is still a fit: the model loads, with less headroom than we like to promise. A struck-through “Won’t fit” means the weights alone overflow the pool. Every model name links to its own page, where each machine that holds it is priced live.

Read the middle rows twice. MiniMax-M2 and Qwen3-235B miss Q4 by twenty-something gigabytes, then load at Q3 — tight, but loaded. The quant dial is worth a full machine tier, and it is the one dial you can still turn after the purchase.

Below the line sit the models this page is really about, so the table exempts them rather than waste a row. DeepSeek-R1 and DeepSeek-V3 store 671B parameters: at Q3 they need more than 300GB, and no 128GB machine pretends otherwise. Kimi-K2-Instruct, Qwen3-Coder-480B, Llama-4-Behemoth and DeepSeek-V4-Pro are further below it still.

These are the models that make 256GB the tier above, not a luxury. If the model you came for is on that list, the honest advice is to decide the model first and let it pick the machine. The memory guide does that arithmetic model-first, if you would rather start there.

Four machines on the Grid carry this much memory, and they are not interchangeable. Memory decides what loads; bandwidth decides how fast it reads. The speed column is Llama-3.3-70B at Q4_K_M, computed the same way on every row.

Same 120GB usable, more than double the tokens per second at the top of the table. That spread is the real cost of buying this class: not whether the model loads, but whether you enjoy using it. The DGX Spark and Ascent GX10 trade speed for an NVIDIA software stack; the refurbished M4 Max trades it for several hundred dollars.

Pick a model and the fit finder ranks every machine that holds it, cheapest first. The Grid lists all of them with prices checked hourly.

The full formulas, the reserve rules and the calibration sources are on the methodology page. A skeptical reader can check the work instead of trusting it.