Is 64 GB of unified memory enough for AI? What 64 GB runs, and why the usable share is closer to 48 GB.
Short answer. Yes for a 27B to 32B model serving a small team, and only just for a 70B model at 4-bit with a short context. The label overstates the budget: AMD says up to 48 GB of a 64 GB Ryzen AI Max+ machine can be set aside as graphics memory, so plan around about 48 GB. A 27B model at 8-bit (about 29 GB) with caches for four long conversations fits inside that. A 70B model at 4-bit (about 42 GB) leaves almost no room for context, and gpt-oss-120b (a 63 GB file) does not fit. Figures are estimates.
01
Start with the usable share, not the label
Unified memory is one pool shared by the processor and the graphics engine. The operating system and the programs need part of it, so the model never gets all 64 GB. For its Ryzen AI Max+ chips AMD describes a setting called Variable Graphics Memory and says a 64 GB machine can dedicate up to 48 GB to graphics, and a 128 GB machine up to 96 GB. Whatever is dedicated to graphics is no longer available as ordinary memory.
Apple’s Metal documentation describes a recommended maximum working set for the GPU rather than the full memory. A community note reports about three quarters of memory by default on Apple silicon, which would also be 48 GB on a 64 GB Mac; that figure is not from Apple, so check the log line of the runtime you use.
Platform
Memory options
Memory bandwidth (vendor)
Note
AMD Ryzen AI Max+ 395 (Strix Halo)
Up to 128 GB LPDDR5x-8000, 256-bit
256 GB/s
Up to 48 GB of 64 GB as graphics memory, per AMD
NVIDIA DGX Spark (GB10)
64 GB (through OEM partners) or 128 GB
273 GB/s
Developer machine; 240 W power supply
Apple Mac Studio, M5 Max
36, 48, 64 or 128 GB
460 GB/s (32-core GPU) or 614 GB/s (40-core GPU)
GPU share set by macOS
02
What fits in a 48 GB budget
Weights use parameters in billions × bits per weight ÷ 8 (Q8_0 about 8.5 bits, Q4_K_M about 4.8) or the published file size; caches use NVIDIA’s per-token formula. Whisper large-v3 is counted at 3.1 GB (its 16-bit file) and bge-m3 at about 1.1 GB (half its 2.27 GB file, my estimate for 16-bit). All estimates.
Case
Weights + caches + 2 GB overhead
Total
Fits in 48 GB?
Qwen3.8-27B at Q8_0, 4 conversations at 32,768 tokens
28.7 + 9.2 + 2.0
39.9 GB
Yes
Qwen3-32B at Q8_0, 2 conversations at 8,192 tokens
34.8 + 4.3 + 2.0
41.1 GB
Yes
Qwen3-32B at Q8_0, 2 conversations at 32,768 tokens
34.8 + 17.2 + 2.0
54.0 GB
No
Llama 3.3 70B at Q4_K_M, 1 conversation at 8,192 tokens
42.0 + 2.7 + 2.0
46.7 GB
Only just
Llama 3.3 70B at Q4_K_M, 1 conversation at 32,768 tokens
A 70B model at 8-bit: about 74 GB for the weights alone.
gpt-oss-120b or anything of 100 billion parameters or more, except at very low precision.
Several large models loaded together while also keeping long conversations for many people.
A dense 32B model at 8-bit with long documents for more than one person.
04
Where this meets ANVE’s plans
ANVE One is specified with 64 GB of unified memory, a 27B-class default model (Qwen 3.8 27B) and a fast 4B model, for 1 to 3 named users. The office-stack row above is that shape of workload. These are launch specifications with indicative prices; no sealed unit exists yet, and ANVE has not published concurrency or speed figures.
05
Questions people ask
Can a 64 GB machine use all 64 GB for the model?
No. The operating system needs memory, and the graphics side is capped: AMD says up to 48 GB on a 64 GB Ryzen AI Max+ machine. Plan the model, its caches and its runtime inside about 48 GB.
Is 64 GB enough for a 70B model?
Only at 4-bit with a short context, about 47 GB for one conversation at 8,192 tokens. There is no room for long documents or a second conversation, so a 27B to 32B model is the more practical choice.
Is 64 GB enough for chat and image generation together?
Usually yes. The office stack in the table above, with caches, comes to about 31 GB, and SDXL’s 16-bit files add about 7 GB. Load the image model only when it is needed.
06
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.
