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 single 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 best fit.
The best local LLM for the Ryzen AI Max+ 395 is Kolibri 1 (est. 47 tok/s, typical error 25%) on its 110GB unified memory. It uses ~49.8GB of unified memory; the Ryzen AI Max+ 395 handles up to 120B parameter models at Q4. (ModelFit, 2026)
Sizing rule: a Q4 model needs about 0.6 GB of unified memory per billion parameters, and ModelFit budgets 90% of this card's 110GB for weights, context, and KV-cache. The per-size table below uses that same budget. Other strong fits: Qwen3.5 122B-A10B Instruct (122B, ~72GB) and Qwen3.6 35B-A3B (Q8) (35B, ~38.7GB). On this card Kolibri 1 runs at est. 47 tok/s, typical error 25%. What it will not run: dense models above ~120B at Q4 exceed this budget and need a lower quant, CPU offload, or a second GPU.
Speeds are ModelFit estimates from memory bandwidth and model size, not measured benchmarks. Typical error: the median gap between the ModelFit speed model and measured runs on hardware the model was not fitted on, for this backend and model size.
Cite this page: ModelFit, AMD Ryzen AI Max+ 395 Local LLM (2026): 128GB Runs 70B-120B, https://modelfit.io/gpu/ryzen-ai-max-395/, updated October 2026, CC BY 4.0.
| Model size | Est. speed | Fit on 110GB |
|---|---|---|
| 7B | est. 37 tok/stypical error 10% | Fits in unified memory |
| 14B | est. 20 tok/stypical error 10% | Fits in unified memory |
| 20B MoE (3.6B active) | est. 43 tok/stypical error 25% | Fits in unified memory |
| 32B | est. 9 tok/stypical error 50% | Fits in unified memory |
| 35B MoE (3B active) | est. 50 tok/stypical error 25% | Fits in unified memory |
| 70B | est. 4 tok/stypical error 10% | Fits in unified memory |
| 120B MoE (5.1B active) | est. 32 tok/stypical error 25% | Fits in unified memory |
Decode speed, generic Q4_K_M model of each size at 4K context, 256 GB/s, typical error 10-50%
ModelFit estimates, not measured benchmarks: decode speed of a generic Q4_K_M model of each size at 4K context on the Ryzen AI Max+ 395's 256 GB/s bandwidth. 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. What typical error means.
Kolibri 1 loads ~49.8 GB of weights; at 16k context the KV cache adds ~0.3 GB (still fits the ~99 GB usable unified memory), and at 64k it adds ~1.3 GB (still fits).
The model leaves about 49 GB of the usable unified memory budget free after weights and 16k context.
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.
One email when a new open-weight model is a better fit for 110 GB of unified memory, plus the Thursday weekly.
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
ModelFit may earn a commission on purchases through these links, at no extra cost to you.
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 est. 4 tok/s, typical error 10%. Mixture-of-Experts models, which activate only a few billion parameters per token, are where this chip performs best, hitting est. 50 tok/s, typical error 25% on a 35B-A3B and est. 32 tok/s on a 120B MoE. 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.
Best for: Reasoning, German, Long context·Pop: 75/100
Perf: est. 47 tok/s, typical error 25% · first token ~0.4s
Fits in 110 GB unified memory with room to spare. Best for reasoning, german, long context on Ryzen AI Max+ 395.
Best for: Frontier-level reasoning, Complex tasks·Pop: 75/100
Perf: est. 23 tok/s, typical error 25% · first token ~0.7s
Fits in 110 GB unified memory with room to spare. Best for frontier-level reasoning, complex tasks on Ryzen AI Max+ 395.
Best for: Reasoning, Coding, Agents·Pop: 88/100
Perf: est. 40 tok/s, typical error 25% · first token ~1.1s
Fits in 110 GB unified memory with room to spare. Best for reasoning, coding, agents on Ryzen AI Max+ 395.
Best for: Agentic coding, Long context, Reasoning·Pop: 55/100
Perf: est. 42 tok/s, typical error 25% · first token ~0.5s
Fits in 110 GB unified memory with room to spare. Best for agentic coding, long context, reasoning on Ryzen AI Max+ 395.
Best for: Agentic coding, Long context, Reasoning·Pop: 55/100
Perf: est. 37 tok/s, typical error 25% · first token ~0.5s
Fits in 110 GB unified memory with room to spare. Best for agentic coding, long context, reasoning on Ryzen AI Max+ 395.
Best for: Chat, Coding, Long Context·Pop: 80/100
Perf: est. 54 tok/s, typical error 25% · first token ~0.4s
Fits in 110 GB unified memory with room to spare. Best for chat, coding, long context on Ryzen AI Max+ 395.
Best for: Agentic coding, Reasoning, Multimodal·Pop: 70/100
Perf: est. 31 tok/s, typical error 25% · first token ~3.4s
Fits in 110 GB unified memory with room to spare. Best for agentic coding, reasoning, multimodal on Ryzen AI Max+ 395.
Best for: Reasoning, Coding, Agents·Pop: 88/100
Perf: est. 36 tok/s, typical error 25% · first token ~2.6s
Fits in 110 GB unified memory with room to spare. Best for reasoning, coding, agents on Ryzen AI Max+ 395.
Best for: Chat, Coding, Long Context·Pop: 80/100
Perf: est. 35 tok/s, typical error 25% · 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.
Best for: Agentic workflows, Long context, Reasoning·Pop: 75/100
Perf: est. 17 tok/s, typical error 25% · first token ~1.0s
Fits in 110 GB unified memory with room to spare. Best for agentic workflows, long context, reasoning on Ryzen AI Max+ 395.
No stated VRAM figure or published run covers AMD for these models yet. By memory size: video models · image models.
Speeds are estimates for Qwen3 8B Q4_K_M, typical error 5%.
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 est. 4 tok/s, typical error 10% because the 256 GB/s memory bandwidth is the bottleneck. MoE models run much faster, est. 50 tok/s, typical error 25% for a 35B-A3B like Qwen3 30B-A3B and est. 32 tok/s for a 120B MoE like gpt-oss-120b, 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.
It launched around $1,999, but that era is over: the 2026 memory shortage repriced 128GB unified-memory boxes sharply upward, and the price now varies widely with storage. Check the live price before budgeting. Note the cheaper EVO-X2 is the 64GB version, which cannot hold a 235B model.
ModelFit estimates a 32B model on the Ryzen AI Max+ 395 at Q4_K_M runs at est. 9 tok/s, typical error 50%. The current 27B-class pick in the catalog is Qwen3.8 27B (ollama run qwen3.8:27b).
Time a model that fits your machine with Ollama, using the same prompts as every other run. The CLI shows your result and asks before it sends anything; your machine then joins the speed test leaderboard. It does not change the quality benchmark.
Use our interactive wizard to compare models across Apple Silicon and NVIDIA GPUs.