Which LLMs Fit in 128 GB of Unified Memory?

Mistral Medium 3.5 128B ranks first at 89.2 GB, with room to spare. It's dense, so every token reads all 128B parameters, and on these machines it writes text slowly.

The 120B mixture of experts models suit this memory best. gpt-oss 120B needs 80.9 GB, Nemotron 3 Super 120B 80.8 GB and Qwen3.5 122B 82.3 GB, and all three still fit at 32K. They run 5B to 12B parameters per token, so they stay quick on the slower memory these machines use.

The table's order comes from the model finder's best pick, which weighs size against release date. Use it to pick a few to try.

Mistral Medium 3.5 128B still fits at 32K, at 98.9 GB. That's past what a 96 GB card holds.

The MoE models leave even more spare. gpt-oss 120B's KV cache takes 1.3 GB at 32K, against 11.9 GB for Mistral Medium 3.5, because half its layers only cache the most recent tokens. For 64K or 128K, open the model finder from the card below the table and change the context there.

At Q8_0, the 70B class fits with room, even at 32K. Llama 3.3 70B needs 93.5 GB there and Qwen 2.5 72B 95.8 GB. The 120B MoE models don't fit at Q8_0. Nemotron 3 Super 120B needs 140.2 GB.

So the choice here is a 120B MoE model at Q4_K_M or a 70B dense model at Q8_0. The MoE models are newer and much faster on this kind of memory, which makes them the usual pick.

Qwen3.8-Flash-Next 180B (MoE) needs 121.2 GB at Q4_K_M and misses by 1.2 GB. At Q3_K_M it needs 97.4 GB and fits, so it's worth trying at the smaller quant.

The 230B class stays out of reach. Qwen3 235B (MoE A22B) needs 128.4 GB even at Q3_K_M. For those, see the 256 GB Mac list.

The calculator gives an M5 Max Mac and the DGX Spark 120 GB each for models. A Ryzen AI Max+ 395 machine gets 112 GB on Windows, where AMD's Variable Graphics Memory setting sets the GPU's share, and Linux can give it more with a kernel setting. Every row here with a total under 112 GB fits the AMD machine too. That includes Mistral Medium 3.5 128B at 32K.

On the Mac, macOS keeps part of the memory away from the GPU by default. On macOS Sonoma and later, this command lets the GPU use 120 GB until the next restart:

sudo sysctl iogpu.wired_limit_mb=122880

All three have much less memory bandwidth than a high end graphics card, so the same model writes tokens more slowly on them. That's why the MoE models suit them: they read only a small part of their weights for each token.