The Ryzen AI Max+ 395 is AMD's answer to Apple Silicon: a single APU with 128GB of unified memory, of which roughly 110GB is GPU-addressable on Linux. That capacity lets a sub-$2,000 mini PC load 70B and even 120B-class models that no consumer GPU can hold. The trade-off is bandwidth: at 256 GB/s, dense large models load but generate slowly, so MoE models in the 30B-120B range are the sweet spot.
The best local LLM for the Ryzen AI Max+ 395 is GPT-OSS 120B at ~11 tok/s on its 110GB unified memory. It uses ~65.4GB of unified memory; the Ryzen AI Max+ 395 handles up to 120B parameter models at Q4.
Speeds are ModelFit estimates from memory bandwidth and model size, not measured benchmarks.
*128GB GMKtec EVO-X2
| Model Size | Est. Speed | Fit on 110GB |
|---|---|---|
| 7B | ~34 tok/s | Fits in unified memory |
| 14B | ~19 tok/s | Fits in unified memory |
| 20B MoE (3.6B active) | ~29 tok/s | Fits in unified memory |
| 32B | ~9 tok/s | Fits in unified memory |
| 35B MoE (3B active) | ~24 tok/s | Fits in unified memory |
| 70B | ~5 tok/s | Fits in unified memory |
| 120B MoE (5.1B active) | ~11 tok/s | Fits in unified memory |
ModelFit estimates, not measured benchmarks: anchored to an 8B-class Q4_K_M model at 16K context on the Ryzen AI Max+ 395's 256 GB/s bandwidth, then scaled by model size. MoE rows scale by active parameters (decode reads only the active experts), so a 35B MoE runs far faster than a dense 32B. "CPU offload" sizes exceed the 110GB unified memory; dense models slow to a crawl there, MoE models degrade less because hot experts stay GPU-resident.
Context costs unified memory too. GPT-OSS 120B loads ~65.4 GB of weights; at 16k context the KV cache adds ~6.0 GB (still fits the ~99 GB usable unified memory), and at 64k it adds ~24.0 GB (still fits).
KV-cache figures assume an fp16 cache, the llama.cpp/Ollama default. Standard GQA models use a size-class estimate (8 KV heads x 128 head dim class); hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next) use the exact per-token cost from their published config, since only their sparse full-attention layers cache KV. A q8_0 KV cache roughly halves either figure. Estimates, not measurements.
A Gen4 M.2 drive keeps your whole GGUF and quant collection on fast local storage, loading models straight off NVMe.
Check price on Amazon40Gbps external storage fast enough to run models from. Pair it with an M.2 drive for a portable model vault.
Check price on AmazonModelFit may earn a commission on purchases through these links, at no extra cost to you. Prices shown are approximate street references.
Unlike a discrete GPU's fixed VRAM, Strix Halo shares one 128GB LPDDR5x pool between CPU and GPU. On Linux, kernel GTT tuning exposes about 110GB of that to the Radeon 8060S iGPU, more than triple an RTX 5090's 32GB. A 70B model at Q4 (~42GB) or a 120B MoE (~65GB) fits with headroom. The catch is memory bandwidth: 256 GB/s (about 215 GB/s measured) is a fraction of a discrete GPU's, and since token generation is bandwidth-bound, dense 70B models run around 5 tok/s. Mixture-of-Experts models, which activate only a few billion parameters per token, are where this chip shines, hitting 50-70+ tok/s. On Windows the GPU is capped at a fixed BIOS allocation with no equivalent shared pool, so the big-model capability is mainly a Linux story today.
GPT-OSS / 117B / MXFP4 / ~65.4 GB
Best for: Reasoning, Coding, Agents·Pop: 88/100
Perf: ~11 tok/s · first token ~1.4s
Fits in 110 GB unified memory with room to spare. Best for reasoning, coding, agents on Ryzen AI Max+ 395.
Qwen / 80B / Q4_K_M / ~50.4 GB
Best for: Chat, Coding, Long Context·Pop: 80/100
Perf: ~17 tok/s · first token ~1.2s
Fits in 110 GB unified memory with room to spare. Best for chat, coding, long context on Ryzen AI Max+ 395.
Qwen / 122B / Q4_K_M / ~72 GB
Best for: Frontier-level reasoning, Complex tasks·Pop: 75/100
Perf: ~9 tok/s · first token ~1.5s
Fits in 110 GB unified memory with room to spare. Best for frontier-level reasoning, complex tasks on Ryzen AI Max+ 395.
Llama / 109B / Q4_K_M / ~67 GB
Best for: Long context, Quality, Multimodal·Pop: 86/100
Perf: ~7 tok/s · first token ~1.6s
Fits in 110 GB unified memory with room to spare. Best for long context, quality, multimodal on Ryzen AI Max+ 395.
Qwen / 35B / Q8_0 / ~38.7 GB
Best for: Reasoning, Coding, Agents·Pop: 88/100
Perf: ~15 tok/s · first token ~1.2s
Fits in 110 GB unified memory with room to spare. Best for reasoning, coding, agents on Ryzen AI Max+ 395.
Qwen / 35B / Q8_0 / ~38.7 GB
Best for: Reasoning, Coding, Agent scenarios·Pop: 90/100
Perf: ~15 tok/s · first token ~1.2s
Fits in 110 GB unified memory with room to spare. Best for reasoning, coding, agent scenarios on Ryzen AI Max+ 395.
Qwen / 27B / Q8_0 / ~30 GB
Best for: Coding, Quality, Long context·Pop: 92/100
Perf: ~7 tok/s · first token ~1.1s
Fits in 110 GB unified memory with room to spare. Best for coding, quality, long context on Ryzen AI Max+ 395.
Qwen / 80B / Q8_0 / ~84.8 GB
Best for: Chat, Coding, Long Context·Pop: 80/100
Perf: ~11 tok/s · first token ~1.4s
Fits in 110 GB unified memory with room to spare. Best for chat, coding, long context on Ryzen AI Max+ 395.
Qwen / 35B / Q4_K_M / ~22 GB
Best for: Reasoning, Coding, Agents·Pop: 88/100
Perf: ~24 tok/s · first token ~1.1s
Fits in 110 GB unified memory with room to spare. Best for reasoning, coding, agents on Ryzen AI Max+ 395.
Qwen / 35B / Q4_K_M / ~20 GB
Best for: Reasoning, Coding, Agent scenarios·Pop: 90/100
Perf: ~24 tok/s · first token ~1.1s
Fits in 110 GB unified memory with room to spare. Best for reasoning, coding, agent scenarios on Ryzen AI Max+ 395.
The Ryzen AI Max+ 395 tops out around up to 120b parameter models. For anything bigger, an hourly rented GPU runs the same open weights with the same Ollama workflow, billed by the hour, no hardware purchase needed.
ModelFit may earn a commission on sign-ups made through these links, at no extra cost to you.
Alibaba Cloud: Widest size range (0.5B to 235B)
LlamaMeta: Most popular open-weight model family
DeepSeekDeepSeek AI: Best-in-class reasoning with R1 models
MistralMistral AI: Excellent performance-per-parameter ratio
GemmaGoogle DeepMind: Excellent quality at small sizes (1B-9B)
PhiMicrosoft: Best quality-per-gigabyte at small sizes
Up to 120B-parameter models. Its 128GB unified memory (~110GB GPU-addressable on Linux) holds a 70B model at Q4 (~42GB) or a 120B MoE (~65GB) with room to spare, far beyond any consumer GPU. Mixture-of-Experts models in the 30B-120B range run best.
It depends on the model type. Dense 70B models generate around 5 tok/s because the 256 GB/s memory bandwidth is the bottleneck. MoE models like Qwen3 30B-A3B or gpt-oss-120b run much faster, 50-70+ tok/s, since only a few billion parameters are active per token. All figures are estimates.
They win on different axes. The RTX 5090 (32GB, 1,792 GB/s) is far faster per token for models that fit in 32GB. The Ryzen AI Max+ 395 (110GB usable, 256 GB/s) is slower but holds models 3x larger. AMD claims up to 3x the 5090-class performance only when a model exceeds the Nvidia card's VRAM and spills to system RAM.
For the largest models, effectively yes. On Linux, GTT kernel tuning lets the GPU address roughly 110GB of the 128GB pool. On Windows the GPU is limited to a fixed BIOS memory carve-out with no equivalent shared pool, so the very-large-model capability is mainly a Linux feature today.
The 128GB GMKtec EVO-X2 launched around $1,999, with street prices roughly $1,800-$2,300. AMD's own first-party dev kit is reported at $3,999. The cheaper $1,499 EVO-X2 is the 64GB version, which cannot hold a 235B model.
ModelFit estimates a 32B model on the Ryzen AI Max+ 395 runs at roughly 9 tok/s at Q4_K_M. The current 27B-class pick in the catalog is Qwen3.5 27B Instruct (ollama run qwen3.5:27b).
New open-weight models, real Apple Silicon benchmarks, and the one model worth running on your Mac this week. Free, one email a week, unsubscribe anytime.
By subscribing you agree to our Privacy Policy and to receive the weekly email. Unsubscribe anytime.
Use our interactive wizard to compare models across Apple Silicon and NVIDIA GPUs.