ModelVRAM
EN 中文

What LLMs can a Ryzen AI Max+ 395 (128 GB) run?

A Ryzen AI Max+ 395 (128 GB), with 96 GB usable by the GPU, runs 48 of the 69 open models tracked here at Q4_K_M with 8K context; the largest is Mistral Medium 3.5 128B (127.7B parameters, 82.7 GB). It reads memory at 256 GB/s; up to 96 GB can be assigned to the GPU. 2 more fit at a lower precision. Each fit leaves at least 0.5 GB free on the Ryzen AI Max+ 395 (128 GB) (why a card does not give a model all its memory).

Models that fit on one Ryzen AI Max+ 395 (128 GB)

The most precise weights that still fit with 8K tokens of context and one request, the memory that takes, the longest context at that precision and the writing speed. Bigger models come first. gpt-oss-120b (67.7 GB, 23–39 tokens/s on the Ryzen AI Max+ 395 (128 GB)) and gpt-oss-20b (14.8 GB, 28–48 tokens/s on the Ryzen AI Max+ 395 (128 GB)) are listed at MXFP4 only, as published.

Model Best precision Memory Longest context Tokens/s
Step 3.7 Flash 196B-A11B 201.4B, 11B active GGUF Q2_K 87.4 GB 164K 15–25
Qwen3.8 Flash Next 180.0B, 6B active GGUF Q3_K_M 90.8 GB 189K 24–40
Mistral Medium 3.5 128B 127.7B GGUF Q4_K_M 82.7 GB 41K 1.8–2.4
Qwen3.5 122B-A10B 125.1B, 10B active GGUF Q5_K_M 91.5 GB 161K 10–17
Nemotron 3.5 Super VL 120B-A12B 124.5B, 12B active GGUF Q5_K_M 90.9 GB 539K 8.8–15
Nemotron 3 Super 120B-A12B 123.6B, 12B active GGUF Q5_K_M 90.3 GB 256K (full) 8.8–15
gpt-oss-120b 116.8B, 5.1B active As published (MXFP4) 67.7 GB 128K (full) 23–39
AliceAI Foundation 80B-A3B 81.3B, 3B active GGUF Q8_0 89.2 GB 252K 22–37
Qwen3-Coder-Next 79.7B, 3B active GGUF Q8_0 87.4 GB 256K (full) 22–37
Kolibri-1 78.1B, 3.5B active GGUF Q8_0 85.8 GB 256K (full) 19–32
Llama 3.1 70B 70.6B GGUF Q8_0 80.0 GB 52K 1.8–2.5
K2-Horizon MoVA 36B-A4B 37.4B, 4B active FP16 / BF16 78.9 GB 88K 7.9–13
Ornith 1.5 35B-A3B 36.0B, 3B active FP16 / BF16 74.3 GB 256K (full) 12–21
Qwen3.6 35B-A3B 36.0B, 3B active FP16 / BF16 74.3 GB 256K (full) 12–21
Ornith 1.0 35B 35.1B, 3B active FP16 / BF16 72.6 GB 256K (full) 12–21
LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active FP16 / BF16 66.9 GB 64K (full) 9.2–16
Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active FP16 / BF16 65.3 GB 256K (full) 11–18
Gemma 4 31B 31.3B FP16 / BF16 66.6 GB 256K (full) 2.2–3.0
GLM-4.7 Flash 31.2B, 3B active FP16 / BF16 64.9 GB 198K (full) 12–20
Xing 4.0 29B-A4B 31.2B, 4B active FP16 / BF16 64.8 GB 256K (full) 9.0–15
Qwen3-Coder 30B-A3B 30.5B, 3.3B active FP16 / BF16 63.9 GB 256K (full) 10–17
Qwen3 30B-A3B 30.5B, 3.3B active FP16 / BF16 63.9 GB 40K (full) 10–17
Muse Glimmer 30B 29.8B FP16 / BF16 61.7 GB 128K (full) 2.3–3.2
Granite 4.2 30B 29.3B FP16 / BF16 62.7 GB 127K 2.3–3.2
ThinkingCap Qwen3.8 27B 27.8B FP16 / BF16 58.0 GB 256K (full) 2.5–3.4
Qwen3.6 27B 27.8B FP16 / BF16 58.0 GB 256K (full) 2.5–3.4
Qwen3.8 27B 27.8B FP16 / BF16 58.0 GB 256K (full) 2.5–3.4
Hemmingway-1 27B 27.3B FP16 / BF16 57.0 GB 256K (full) 2.5–3.5
Gemma 4 26B-A4B 25.8B, 4B active FP16 / BF16 53.9 GB 256K (full) 8.9–15
gpt-oss-20b 20.9B, 3.6B active As published (MXFP4) 14.8 GB 128K (full) 28–48
Gemma 4 12B 12.0B FP16 / BF16 25.7 GB 256K (full) 5.7–7.8
ZDTaichu 5.0 9B 9.8B FP16 / BF16 20.8 GB 128K (full) 7.0–9.6
Ornith 1.5 9B 9.7B FP16 / BF16 20.6 GB 256K (full) 7.1–9.8
Qwen3.5 9B 9.7B FP16 / BF16 20.6 GB 256K (full) 7.1–9.8
LensVLM 9B 9.4B FP16 / BF16 20.1 GB 256K (full) 7.3–10
Ornith 1.0 9B 9.4B FP16 / BF16 20.1 GB 256K (full) 7.3–10
MiMo V2.6 Distill Qwen 9B 9.4B FP16 / BF16 20.1 GB 256K (full) 7.3–10
Bespoke Nimble 9B 9.0B FP16 / BF16 19.1 GB 8K (full) 7.7–11
Granite 4.2 8B 8.8B FP16 / BF16 19.9 GB 128K (full) 7.4–10
LFM2.5 8B-A1B 8.5B, 1.5B active FP16 / BF16 18.0 GB 128,000 (full) 24–40
Qwen3 8B 8.2B FP16 / BF16 18.5 GB 40K (full) 7.9–11
Llama 3.1 8B 8.0B FP16 / BF16 18.1 GB 128K (full) 8.1–11
Gemma 4 E4B 8.0B FP16 / BF16 17.1 GB 128K (full) 8.6–12
Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active FP16 / BF16 16.7 GB 128K (full) 28–47
LightOnOCR-3 4B 4.5B FP16 / BF16 10.1 GB 256K (full) 15–20
Spark-X2.5 4B 4.1B FP16 / BF16 9.35 GB 1M (full) 16–22
Nemotron 3 Nano 4B 4.0B FP16 / BF16 8.78 GB 256K (full) 17–23
Granite 4.2 3B 3.7B FP16 / BF16 8.69 GB 128K (full) 17–24
MiniCPM5 2B 2.5B FP16 / BF16 6.02 GB 128K (full) 25–35
Limite 1B Violetto 1.0B FP16 / BF16 2.79 GB 128K (full) 58–82

Too big for one Ryzen AI Max+ 395 (128 GB) at 8K context

How many Ryzen AI Max+ 395 (128 GB) cards these need at Q4_K_M with 8K context and 0.5 GB to spare, in one tensor-parallel group; some need more than 8.

Best models for Ryzen AI Max+ 395 (128 GB) with room for long context

The biggest models that still fit at Q4_K_M with 32,768 tokens of context (or their whole window, if shorter), the memory left over, the longest context the card takes at that precision and the writing speed with that context in the cache, from 256 GB/s of bandwidth. gpt-oss-120b is at MXFP4, as published.

ModelContextMemoryLeft overLongest contextTokens/s
Mistral Medium 3.5 128B 127.7B 32K 91.8 GB 4.25 GB 41K 1.6–2.2
Qwen3.5 122B-A10B 125.1B 32K 78.9 GB 17.1 GB 256K (full) 11–18
Nemotron 3.5 Super VL 120B-A12B 124.5B 32K 77.9 GB 18.1 GB 1M (full) 10–17
Nemotron 3 Super 120B-A12B 123.6B 32K 77.4 GB 18.6 GB 256K (full) 10–17
gpt-oss-120b 116.8B, MXFP4 32K 68.6 GB 27.4 GB 128K (full) 18–31

llama-server commands for one Ryzen AI Max+ 395 (128 GB)

GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on the Ryzen AI Max+ 395 (128 GB). No -ngl: llama.cpp’s -fit, on by default since b7440, places the layers.

Mistral Medium 3.5 128B, Q4_K_M with 41K tokens: 1.5–2.1 tokens/s on the Ryzen AI Max+ 395 (128 GB)

llama-server -hf unsloth/Mistral-Medium-3.5-128B-GGUF:Q4_K_M -c 41984

Q4_K_M/Mistral-Medium-3.5-128B-Q4_K_M-00001-of-00003.gguf, 74.9 GB: 95.2 GB used and 867 MB free of the Ryzen AI Max+ 395 (128 GB)'s 96 GB usable, 1.5–2.1 tokens/s once the 41K cache is full.

Qwen3.5 122B-A10B, Q4_K_M with 256K tokens: 6.1–10 tokens/s on the Ryzen AI Max+ 395 (128 GB)

llama-server -hf unsloth/Qwen3.5-122B-A10B-GGUF:Q4_K_M -c 262144

Q4_K_M/Qwen3.5-122B-A10B-Q4_K_M-00001-of-00003.gguf, 76.5 GB: 85.5 GB used and 10.5 GB free of the Ryzen AI Max+ 395 (128 GB)'s 96 GB usable, 6.1–10 tokens/s once the 256K cache is full.

How fast the Ryzen AI Max+ 395 (128 GB) writes as the context fills

Tokens per second at Q4_K_M (gpt-oss-120b at MXFP4) for the biggest models that fit with 32K tokens, with 8K and 32K tokens in the cache and at the longest context the card holds, and at Q8_0 with 8K where that fits. Each token reads the active weights and the whole cache once, so the speed follows the 256 GB/s of bandwidth and falls as the cache grows. The ceiling is that bandwidth divided by the bytes read per token, which no runtime reaches; the reply time is for 1,000 new tokens with 32K (or the longest context) in the cache, prompt processing aside. With the same memory, M4 Max Mac (128 GB), 546 GB/s, writes 113% faster by bandwidth; RTX PRO 6000 Blackwell, 1,792 GB/s, writes 600% faster by bandwidth; 4× RTX 3090, 3,744 GB/s, writes 1363% faster by bandwidth.

Model8K32KLongestAt the longestQ8_0 at 8KBandwidth ceiling at 8K1,000-token reply at 32K
Mistral Medium 3.5 128B 1.8–2.4 1.6–2.2 41K 1.5–2.1 Does not fit 3 464–634 s
Qwen3.5 122B-A10B 12–20 11–18 256K 6.1–10 Does not fit 41 54–91 s
Nemotron 3.5 Super VL 120B-A12B 10–17 10–17 1M 4.8–8.0 Does not fit 35 59–100 s
Nemotron 3 Super 120B-A12B 10–17 10–17 256K 8.1–14 Does not fit 35 59–100 s
gpt-oss-120b 23–39 18–31 128K 9.8–16 MXFP4 only 81 32–55 s
AliceAI Foundation 80B-A3B 36–62 28–48 256K 9.2–15 22–37 127 21–36 s
Qwen3-Coder-Next 36–62 28–48 256K 9.2–15 22–37 127 21–36 s
Kolibri-1 31–53 26–44 256K 10.0–17 19–32 108 23–39 s
Llama 3.1 70B 3.1–4.2 2.6–3.6 128K 1.6–2.2 1.8–2.5 6 279–381 s
K2-Horizon MoVA 36B-A4B 19–31 8.6–14 348K 1.1–1.8 13–22 64 70–117 s

Near misses on Ryzen AI Max+ 395 (128 GB)

Models that miss at Q4_K_M with 32,768 tokens of context (or their whole window), or leave under 0.5 GB free there, but fit with a shorter context or 3-bit weights, or are short by at most a quarter of the memory and fit with MoE offload. Smallest shortfall first.

ModelShort by at Q4_K_MFits instead
Qwen3.8 Flash Next 180.0B 16.9 GB GGUF Q3_K_M with 32K, 91.5 GB
Step 3.7 Flash 196B-A11B 201.4B 31.1 GB GGUF IQ3_XXS with 32K, 87.4 GB

Run locally or rent a Ryzen AI Max+ 395 (128 GB)?

getdeploying.com lists no on-demand rental of Ryzen AI Max+ 395 (128 GB) (checked ). Among the GPUs this site tracks, the cheapest to rent with at least 96 GB is the RTX PRO 6000 Blackwell at a median $2.19 an hour, which holds 48 of the 69 models at Q4_K_M with 8K context against 48 here; every $1,000 equals about 457 hours of it.

How these numbers are worked out

Memory is the weights at each precision plus the KV cache for 8,192 tokens in FP16 and runtime overhead (0.5 GB plus 10%), computed from each model's files on Hugging Face with the LLM VRAM Calculator. Speed is estimated from memory bandwidth and the parameters read per token with the LLM Speed Calculator. Both tools take any other model from Hugging Face.

Nearby GPUs and setups

How many of the tracked models each one holds at Q4_K_M with 8K context, against 48 here.

SetupHow it relatesMemoryBandwidthModels at Q4_K_M
M4 Max Mac (128 GB) Same memory 96 GB usable 546 GB/s 48 (same)
RTX PRO 6000 Blackwell Same memory 96 GB 1,792 GB/s 48 (same)
4× RTX 3090 Same memory 96 GB 3,744 GB/s 48 (same)
H100 SXM Next size down 80 GB 3,350 GB/s 47 (−1)
DGX Spark (128 GB) Next size up 120 GB usable 273 GB/s 49 (+1)

Model numbers read from Hugging Face on .