AMD Ryzen AI Max+ 395 (Strix Halo) 128GB vs NVIDIA DGX Spark 128GB for local LLMs
Memory fit, speed and context compared with the same math as the advisor. Both offer about the same usable memory (108.3 GB), so the difference is mostly bandwidth (256 vs 273 GB/s).
Verdict from the numbers
- Fit. AMD Ryzen AI Max+ 395 (Strix Halo) 128GB fits 13 of 13 models fully at 16k context (largest: gpt-oss 120B Q8_0); NVIDIA DGX Spark 128GB fits 13 (largest: gpt-oss 120B Q8_0). Both fit the same set of models.
- Speed. On the 13 models that fit fully on both, NVIDIA DGX Spark 128GB is faster by a typical 1.1x. AMD Ryzen AI Max+ 395 (Strix Halo) 128GB is ahead on 4, NVIDIA DGX Spark 128GB on 9. 1 of those comparisons use measured speeds on both sides; the rest involve at least one estimate (see the badges).
- Trade-off. Both offer about the same usable memory (108.3 GB), so the difference is mostly bandwidth (256 vs 273 GB/s).
Where to check prices
Price tier note
We do not list prices because they move daily and differ between new and used stock. What the data does show is what the memory difference buys: both have the same usable memory, so model fit is identical and the choice comes down to speed (256 vs 273 GB/s bandwidth). Use the Amazon links above to compare current listings and decide whether the difference is worth it for the models you want to run.
Model by model · 16k context, f16 KV cache
| Model | AMD Ryzen AI Max+ 395 (Strix Halo) 128GB fit | tok/s | NVIDIA DGX Spark 128GB fit | tok/s | Faster / fits |
|---|---|---|---|---|---|
| Gemma 4 26B-A4B Q8_0 | fits fully | 35 scaled from measured | fits fully | 53 scaled from measured | NVIDIA DGX Spark 128GB |
| Gemma 4 31B Q4_K_M | fits fully | 10 measured | fits fully | 12 estimated | NVIDIA DGX Spark 128GB |
| Gemma 4 12B Q4_K_M | fits fully | 15 scaled from measured | fits fully | 17 scaled from measured | NVIDIA DGX Spark 128GB |
| Qwen3.6 27B Q4_K_M | fits fully | 13 measured | fits fully | 13 estimated | tie |
| Qwen3.6 35B-A3B Q8_0 | fits fully | 53 measured | fits fully | 49 measured | AMD Ryzen AI Max+ 395 (Strix Halo) 128GB |
| Qwen3.5 9B Q5_K_M | fits fully | 16 scaled from measured | fits fully | 17 scaled from measured | NVIDIA DGX Spark 128GB |
| Qwen3.5 4B Q8_0 | fits fully | 23 scaled from measured | fits fully | 50 scaled from measured | NVIDIA DGX Spark 128GB |
| gpt-oss 20B Q8_0 | fits fully | 93 scaled from measured | fits fully | 86 scaled from measured | AMD Ryzen AI Max+ 395 (Strix Halo) 128GB |
| gpt-oss 120B Q8_0 | fits fully | 44 scaled from measured | fits fully | 57 scaled from measured | NVIDIA DGX Spark 128GB |
| Qwen3-Coder 30B-A3B Q8_0 | fits fully | 55 scaled from measured | fits fully | 61 measured | NVIDIA DGX Spark 128GB |
| Mistral Small 3.2 24B Q4_K_M | fits fully | 14 scaled from measured | fits fully | 11 scaled from measured | AMD Ryzen AI Max+ 395 (Strix Halo) 128GB |
| Qwen3-VL 8B Q6_K | fits fully | 16 scaled from measured | fits fully | 19 scaled from measured | NVIDIA DGX Spark 128GB |
| GLM-4.5-Air Q6_K | fits fully | 16 scaled from measured | fits fully | 17 scaled from measured | NVIDIA DGX Spark 128GB |
Badges: measured medians of real llama.cpp runs; scaled from measured and estimated come from our bandwidth model. See sources and methods.
How much context fits?
Largest context where weights plus KV cache stay inside usable memory, per model that fits fully on at least one side. q8_0 KV roughly halves the cache.
| Model | AMD Ryzen AI Max+ 395 (Strix Halo) 128GB f16 | NVIDIA DGX Spark 128GB f16 | AMD Ryzen AI Max+ 395 (Strix Halo) 128GB q8_0 | NVIDIA DGX Spark 128GB q8_0 |
|---|---|---|---|---|
| gpt-oss 120B | 128k | 128k | 128k | 128k |
| GLM-4.5-Air | 32k | 128k | 64k | 128k |
| Qwen3.6 35B-A3B | 1M | 1M | 1M | 1M |
| Gemma 4 31B | 256k | 256k | 256k | 256k |
| Qwen3-Coder 30B-A3B | 512k | 512k | 1M | 1M |
| Qwen3.6 27B | 1M | 1M | 1M | 1M |
| Gemma 4 26B-A4B | 256k | 256k | 256k | 256k |
| Mistral Small 3.2 24B | 128k | 128k | 128k | 128k |
| gpt-oss 20B | 128k | 128k | 128k | 128k |
| Gemma 4 12B | 256k | 256k | 256k | 256k |
| Qwen3.5 9B | 1M | 1M | 1M | 1M |
| Qwen3-VL 8B | 512k | 512k | 1M | 1M |
| Qwen3.5 4B | 1M | 1M | 1M | 1M |
Spec summary
| AMD Ryzen AI Max+ 395 (Strix Halo) 128GB | NVIDIA DGX Spark 128GB | |
|---|---|---|
| Usable memory for model + KV cache | 108.3 GB | 108.3 GB |
| Memory bandwidth | 256 GB/s | 273 GB/s |
| Models that fit fully at 16k | 13 of 13 | 13 of 13 |
| Models that need offload or do not fit | 0 | 0 |
| Measured data points on file | 26 | 19 |
FAQ
Which is faster for local LLMs, AMD Ryzen AI Max+ 395 (Strix Halo) 128GB or NVIDIA DGX Spark 128GB?
On the 13 models that fit fully on both, NVIDIA DGX Spark 128GB is typically 1.1x faster (geometric mean). AMD Ryzen AI Max+ 395 (Strix Halo) 128GB wins 4 and NVIDIA DGX Spark 128GB wins 9. 1 comparisons are measured on both.
Which fits more models, AMD Ryzen AI Max+ 395 (Strix Halo) 128GB or NVIDIA DGX Spark 128GB?
AMD Ryzen AI Max+ 395 (Strix Halo) 128GB fits 13 of 13 models fully at 16k context and NVIDIA DGX Spark 128GB fits 13.
Can I run long context on AMD Ryzen AI Max+ 395 (Strix Halo) 128GB or NVIDIA DGX Spark 128GB?
At f16 KV cache, the largest-model context limits are in the table above (for example gpt-oss 120B on AMD Ryzen AI Max+ 395 (Strix Halo) 128GB: 128k; gpt-oss 120B on NVIDIA DGX Spark 128GB: 128k). A q8_0 KV cache extends it further.
Do you state prices for AMD Ryzen AI Max+ 395 (Strix Halo) 128GB or NVIDIA DGX Spark 128GB?
No. Prices change daily, so we link to Amazon search results instead and focus on measured and estimated speed and memory fit.
Related
AMD Ryzen AI Max+ 395 (Strix Halo) 128GB · NVIDIA DGX Spark 128GB · Measured model pages: gpt-oss 120B · gpt-oss 120B
More comparisons: AMD Ryzen AI Max+ 395 (Strix Halo) 128GB vs MacBook Pro 14/16 (2024) M4 Max 128GB · all comparisons
By memory: 8 GB · 12 GB · 16 GB · 24 GB · 32 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB