Sunk Cost sunkcost.ai Data checked 2026-09-03

What can you run with 96 GB of memory?

A model does not get the 96 GB. On the Mac Studio M5 Ultra, 96GB it gets 72 GB and on the RTX PRO 6000 Blackwell, 96GB it gets 95 GB, because the system keeps a share of one and the card keeps a margin free on the other. That is 29 of the 39 current models at 32k of context on the first and 33 on the second, and the strongest of them is Qwen3.8 27B.

What a model gets72 GB to 95 GB of the 96 GB, depending on the machine. Each machine's page says where the rest goes.
Models that fit at 32k29 to 33 of the 39 current ones, at the quantisation this site lists each at.
The strongest of themQwen3.8 27B, sonnet-class.
Cheapest machine at this sizeMac Studio M5 Ultra, 96GB at $5,499.

Run the numbers on Qwen3.8 27B at 96 GB

Jump to: The machines sold with 96 GB · Models that fit in 96 GB · What 128 GB adds over 96 GB · How much context 96 GB leaves room for · Does a 96 GB machine pay for itself?

The machines sold with 96 GB

There are three machines here with 96 GB, and what they hand a model is not the same figure. The fourth column is what each one holds at 32k of context, and it is the number that machine's own page prints.

MachineWhat a model getsPriceModels it holds at 32kStrongest of them
RTX PRO 6000 Blackwell, 96GB 95 GB $18,000card only 33 Qwen3.8 27B Run the numbers
Mac Studio M3 Ultra, 96GBprevious 72 GB $3,999 29 Qwen3.8 27B Run the numbers
Mac Studio M5 Ultra, 96GB 72 GB $5,499 29 Qwen3.8 27B Run the numbers

List prices, and the memory each maker publishes. A machine marked previous is one that is no longer sold, priced at what it launched at. The RTX PRO 6000 Blackwell, 96GB is priced as the card alone, so add the PC around it before comparing it with a complete computer. All seven cards here are ranked by what each one holds.

Models that fit in 96 GB

Every current model the roomiest 96 GB machine here holds at 32k of context, strongest first. The third column is the whole job at once: the weights plus the cache for 32k of context. The fourth says how many of the machines at this size have room for it, which is the part a size on its own cannot tell you. Speeds are on the RTX PRO 6000 Blackwell, 96GB, the machine this list is cut from; speed follows memory bandwidth rather than memory size, so another machine at 96 GB runs the same model at its own rate, and its page gives every one of them.

ModelParametersNeeds at 32kMachines at 96 GBSpeed on the RTX PRO 6000 Blackwell, 96GB
Qwen3.8 27BQ4_K_M 27.8B 18.6 GB 3 of 3 78 tok/s estimated Run the numbers
Ling 3.0 flashQ4_K_M 124B 81.6 GB 1 of 3 80 tok/s estimated Run the numbers
Qwen3.6 27BQ4_K_M 27.8B 19 GB 3 of 3 77 tok/s estimated Run the numbers
Qwen3.6 35B-A3BQ4_K_M 36B 22.8 GB 3 of 3 221 tok/s estimated Run the numbers
Muse Glimmer 30BQ4_K_M 29.8B 17.8 GB 3 of 3 81 tok/s estimated Run the numbers
Gemma 4 26B-A4BQ4_K_M 25.2B 17.8 GB 3 of 3 162 tok/s estimated Run the numbers
Qwen3.5 122B-A10BUD-Q4_K_M 125.1B 79.1 GB 1 of 3 79 tok/s estimated Run the numbers
Granite 4.2 30BQ4_K_M 29.3B 26.3 GB 3 of 3 55 tok/s estimated Run the numbers
Gemma 4 31B itQ4_K_M 31.3B 25.8 GB 3 of 3 56 tok/s estimated Run the numbers
GLM-4.7-FlashQ4_K_M 31.2B 20.1 GB 3 of 3 157 tok/s estimated Run the numbers
Nemotron 3.5 Lightning 30B-A3BQ4_K_M 31.6B 25.7 GB 3 of 3 212 tok/s estimated Run the numbers
Gemma 4 12BQ4_K_M 12B 8 GB 3 of 3 182 tok/s estimated Run the numbers
Qwen3.5 9BQ4_K_M 9.7B 6.8 GB 3 of 3 215 tok/s estimated Run the numbers
Qwen3.5 4BQ4_K_M 4.7B 3.8 GB 3 of 3 381 tok/s estimated Run the numbers
MiniCPM5 2BQ4_K_M 2.5B 3 GB 3 of 3 489 tok/s estimated Run the numbers
gpt-oss-120bMXFP4 116.8B 64.6 GB 3 of 3 136 tok/s measured Run the numbers
Granite 4.2 8BQ4_K_M 8B 10.7 GB 3 of 3 135 tok/s estimated Run the numbers
Ling 3.0 tinyQ4_K_M 7.9B 6.4 GB 3 of 3 231 tok/s estimated Run the numbers
Mistral Small 4 (119B-2603)Q4_K_M 119.4B 74.5 GB 1 of 3 124 tok/s estimated Run the numbers
Qwen3-Coder NextQ4_K_M 79.7B 49.2 GB 3 of 3 211 tok/s estimated Run the numbers
Qwen3-Coder 30B-A3BQ4_K_M 30.5B 21.8 GB 3 of 3 106 tok/s estimated Run the numbers
Devstral 2 123BQ4_K_M 125B 86.7 GB 1 of 3 17 tok/s estimated Run the numbers
gpt-oss-20bMXFP4 20.9B 12.9 GB 3 of 3 207 tok/s measured Run the numbers
Gemma 4 E4BQAT Q4_0 8B 5.7 GB 3 of 3 161 tok/s estimated Run the numbers
Devstral Small 2 24BQ4_K_M 24B 19.7 GB 3 of 3 74 tok/s estimated Run the numbers
LFM2.5 2.6BQ4_K_M 2.7B 2.2 GB 3 of 3 658 tok/s estimated Run the numbers
Ministral 3 14BQ4_K_M 14B 13.6 GB 3 of 3 107 tok/s estimated Run the numbers
Ministral 3 8BQ4_K_M 8.9B 9.8 GB 3 of 3 149 tok/s estimated Run the numbers
Ornith 1.5 35B-A3BQ4_K_M 36B 22.4 GB 3 of 3 224 tok/s estimated Run the numbers
KAT-Coder V2.5 Dev 35B-A3BQ4_K_M 34.7B 22.1 GB 3 of 3 220 tok/s estimated Run the numbers
Laguna XS 2.1Q4_K_M 33.4B 21.7 GB 3 of 3 172 tok/s estimated Run the numbers
Ornith 1.5 9BQ4_K_M 9.7B 6.9 GB 3 of 3 212 tok/s estimated Run the numbers
Spark-X2.5 4BQ4_K_M 4.1B 3.9 GB 3 of 3 376 tok/s estimated Run the numbers

For scale, the calculator starts from 80 tok/s for a hosted API and times a local machine against it. 26 of the 33 above reach it, and the slowest is 17 tok/s.

Superseded models are left out of the count, because what a machine is worth buying for is what you would run on it today; the leaderboard ranks every model this site lists, older ones included. Where the cache figure comes from is one section of the memory guide.

What 128 GB adds over 96 GB

The machine here with 128 GB that hands a model the most of it is the DGX Spark, 128GB, at 119.5 GB, and it holds 33 of the 39. At 96 GB the most is 95 GB, on the RTX PRO 6000 Blackwell, 96GB, holding 33. Nothing new fits. What the step buys is a longer window and room to work, not a model that was out of reach. The strongest either way is Qwen3.8 27B.

The cheapest machine at 128 GB is the Strix Halo Framework Desktop, 128GB at $3,449. The cheapest at 96 GB is the Mac Studio M5 Ultra, 96GB at $5,499, so the larger size is the cheaper of the two here.

Both sides are read off the machine at each size that hands a model the most, so the comparison is the best case against the best case. The table above has every machine at 96 GB and what each of them holds.

How much context 96 GB leaves room for

The cache grows with the window you ask for, so the same machine holds fewer models the longer the context. One column here for each amount a 96 GB machine hands over. A model is counted only where its own context ceiling reaches that far.

Context72 GB to a modelMac Studio M5 Ultra, 96GB95 GB to a modelRTX PRO 6000 Blackwell, 96GB
4k 2933
8k 2933
16k 2933
32k default 2933
64k 2932
128k 2932

Counted over the 39 current models, at the quantisation each is listed at and with the cache at 16 bits. The calculator has a switch for a smaller cache, which changes both what fits and how fast a long window runs.

Does a 96 GB machine pay for itself?

On the Mac Studio M5 Ultra, 96GB at $5,499, the cheapest machine at 96 GB with a published price, the model that pays it back soonest at 500k tokens a day is Qwen3.8 27B, in 68 years. At 20M tokens a day, agents running most of the day, it is Qwen3.8 27B in 21 months.

The sum is the same everywhere on this site: what the same work costs to rent, less what the electricity costs to generate it, against the price of the machine. The best buys rank the quickest pay-back at every level of use, and what it costs a month puts the machine and the API bill in the same shape.

Put your own usage in

Every figure is at 32k of context unless the row says otherwise, with the cache at 16 bits and each model at the quantisation this site lists it at. Weights are the published file sizes on each model's page, and the cache is worked out from the architecture recorded there. What a model gets is the memory the GPU can address, which each machine's page explains. Speeds say whether anybody measured them; where they were not, they are worked out from memory bandwidth. To change the context, the quantisation or the price you would pay, open the calculator. For the sizes either side of this one, the memory guide has the ladder in full.