What can 128 GB of memory run? The largest open models that fit, and the limits behind the label.
Short answer. A machine with 128 GB of memory can run a 70B model at 8-bit (about 75 GB of weights), OpenAI’s gpt-oss-120b (a 63 GB file), a 32B model at full BF16 precision, or a 27B model at BF16 with its full 262,144-token context. Plan around about 96 GB, not 128: AMD says up to 96 GB of a 128 GB Strix Halo machine can be graphics memory. Models of 100 billion parameters or more fit only at about 4-bit, and the largest open models, such as a 405B model at 249 GB, do not fit at all. Figures are estimates.
01
Plan around 96 GB, not 128 GB
On a Ryzen AI Max+ 395 machine with 128 GB, AMD documents 96 GB as the most that can be dedicated to graphics on Windows, and says the graphics engine can technically reach 112 GB of addressable memory in total. Whatever is dedicated to graphics is taken from the processor’s share. A Mac Studio with 128 GB and the DGX Spark are the other 128 GB machines in this class; macOS caps the GPU’s share by default, about three quarters according to a community note, which is also 96 GB. Each number is a planning figure, not a promise.
02
What fits in about 96 GB
Weights use parameters in billions × bits per weight ÷ 8, or published file sizes where named; caches use 2 × layers × key/value heads × head size × bytes per token. All figures are estimates.
Case
Weights
Cache
Total with 2 GB overhead
Fits?
Llama 3.3 70B at Q8_0, 1 conversation at 32,768 tokens
75.0 GB (file)
10.7 GB
87.7 GB
Yes
Llama 3.3 70B at BF16
About 140 GB
n/a
More than 140 GB
No
gpt-oss-120b (MXFP4), 1 conversation at 131,072 tokens
63.4 GB (file)
9.7 GB (upper bound)
75.1 GB
Yes
Qwen3-32B at BF16, 1 conversation at 32,768 tokens
65.5 GB (file)
8.6 GB
76.1 GB
Yes
Qwen3.8-27B at BF16, 1 conversation at 262,144 tokens
54.0 GB
17.2 GB + 0.15 GB state
73.3 GB
Yes
Llama 4 Scout, 109B with 17B active, at Q4_K_M
About 96 GB, per AMD
Extra
More than 96 GB
Borderline
Llama 3.1 405B at Q4_K_M
249.1 GB (llama.cpp documentation)
Extra
More than 250 GB
No
03
Several models at once
An office rarely runs one model. A realistic stack is a 27B model at Q8_0 (28.7 GB), a 4B model at Q4_K_M (2.4 GB), Whisper large-v3 for speech (3.1 GB), an embedding model such as bge-m3 (about 1.1 GB) and an SDXL image model (about 7.1 GB), with caches for four long conversations on the 27B model and one short one on the 4B model (10.4 GB). With 2 GB of overhead that comes to about 55 GB, leaving about 41 GB of a 96 GB budget for more conversations or a second large model.
04
Three 128 GB machines compared
Machine
Memory
Memory bandwidth (vendor)
Power facts (vendor)
AMD Ryzen AI Max+ 395 (Strix Halo)
Up to 128 GB LPDDR5x-8000
256 GB/s
Configurable chip power of 45 to 120 W
NVIDIA DGX Spark
128 GB
273 GB/s
240 W power supply; GB10 chip rated 140 W
Apple Mac Studio, M5 Max with 40-core GPU
128 GB
614 GB/s
480 W maximum continuous; Apple measured up to 200 W for its M5 Max test unit
05
Where this meets ANVE’s plans
ANVE Office is not a 128 GB machine: at launch it pairs a 24 GB NVIDIA RTX PRO 4000 Blackwell SFF Edition with 64 GB of system memory, for up to 25 named users. ANVE Office Pro is specified with an NVIDIA RTX PRO 4000 Blackwell with 24 GB and 128 GB of system memory, which is a different thing: the model has to fit in the card’s 24 GB to run fast, and the system memory holds the rest of the stack. Launch specifications, indicative prices, no sealed unit yet.
06
Questions people ask
Is 128 GB of ordinary RAM the same as 128 GB of unified memory?
No. With a separate graphics card, a model must fit in the card’s own memory to be fully accelerated, and llama.cpp only partly accelerates a model larger than the card. Unified memory lets the graphics engine use a large share of the same pool.
Can 128 GB run the biggest open models?
No. llama.cpp’s documentation lists a Llama 3.1 405B model at 249.1 GB even at 4-bit. 128 GB suits models up to about 120 billion parameters.
Do I need 128 GB for an office?
Only if you want larger models or many long conversations. A 4-bit 27B model fits comfortably in 64 GB; 128 GB buys 70B-class models at 8-bit, several models loaded together and more simultaneous long documents.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
