GEEKOM A9 Mega (Ryzen AI Max+ 395) 128GB
Key facts
The GEEKOM A9 Mega (Ryzen AI Max+ 395) 128GB has 128 GB shared memory. 35 of 42 models we track fit fully at standard quality (estimate).
€3,999. Checked . GEEKOM (Germany).
GEEKOM's AMD mini PC with 128 GB of shared memory. The graphics get a fixed part of it.
Price
$3,700€3,999
Memory
128 GB shared memory
96 GB of it for the graphics
Good for
Up to Qwen2.5 32B
The largest well-known model it runs well: a mid-size model, at standard quality.
Comfortable~9.4 tok/sSelf-reported speeds on 128 GB Strix Halo
@c4software runs two 128 GB Ryzen AI Max+ 395 computers. bigchuck uses llama.cpp in a ROCm container for Qwen3.8-27B, Qwen3.8-Flash-Next and DeepSeek V4 Flash, mostly at UD-Q4_K_XL, with speculative decoding where it pays. silentchuck is a Minisforum MS-S1 MAX running gufo, his Strix-specific HIP engine. That Minisforum box is a different listing from this one. Since gufo 0.11.0 his gufo measurements run on silentchuck, a Minisforum MS-S1 MAX with a Ryzen AI Max+ 395 and 128 GB, not on bigchuck.
Agentic decode of 50.7 tokens per second was gufo 0.7.1.
Prefill of 1,424 tokens per second at 52,000 tokens of context was gufo 0.8.1.
Those are different gufo versions, not one run. This is self-reported, not a localLLMrig measurement.
Framework Desktop Strix Halo benchmark (self-reported)
Valérian de Thézan de Gaussan measured llama.cpp on a Framework Desktop with an AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU, and 128 GB shared LPDDR5X (~111.56 GiB exposed to Vulkan). The 28 October 2026 follow-up used Arch Linux, Mesa RADV 26.2.5-arch1.1, llama.cpp commit f5222f25c, Vulkan with KHR cooperative matrices, and llama-bench -p 512 -n 128 -t 6 -ngl 999 -r 5. Mistral-Small-4-119B-2603-UD-Q4_K_M (119B total / 6.5B active, 68.70 GiB) prefilled at 675.07 ± 2.995 tok/s and decoded at 32.72 ± 0.020 tok/s. Self-reported, community measurement, not ours.
That Framework Desktop is a different listing from this one. Read Valérian de Thézan de Gaussan's interview.
Check the price
Prices change often. The shop shows today's price.
EU shops · Shown to visitors in Europe.
US shops · Shown to visitors outside Europe.
If you buy through these links we may earn a commission. It doesn't change our ranking. How shop links work
As an Amazon Associate I earn from qualifying purchases.
Als Amazon-Partner verdiene ich an qualifizierten Verkäufen.
Als Amazon-Partner verdien ik aan in aanmerking komende aankopen.
We never pass off an estimate as a benchmark. Commissions never change which hardware we recommend. How we calculate
Which models run on it
Each AI model needs a certain amount of memory. This is what fits at standard quality with a normal-length conversation (about 12 pages of text), and how fast the answers would come.
| AI model | Runs? | Speed |
|---|---|---|
| LFM3-1.7Bsmall | Runs | Fast~164 tok/s |
| Llama 3.2 3Bsmall | Runs | Fast~70 tok/s |
| Spark-X2.6-4Bsmall | Runs | Fast~68 tok/s |
| Mistral 7B v0.3small | Runs | Fast~37 tok/s |
| NVIDIA Nemotron 3.5 Bolt E4Bsmall | Runs | Fast~32 tok/s |
| Llama 3.1 8Bsmall | Runs | Fast~35 tok/s |
| Gemma 2 9Bsmall | Runs | Fast~27 tok/s |
| Qwen2.5 14Bmid-size | Runs | Comfortable~19.6 tok/s |
| Phi-4 14Bmid-size | Runs | Comfortable~19.4 tok/s |
| DeepSeek R1 Distill Qwen 14Bmid-size | Runs | Comfortable~19.2 tok/s |
| gpt-oss-20bmid-size | Runs | Fast~88 tok/s |
| Mistral Small 3.2 24Bmid-size | Runs | Comfortable~13.0 tok/s |
| Gemma 4 26B A4Bmid-size | Runs | Not estimated |
| OrcaCyber Nano 1.0mid-size | Runs | Comfortable~12.6 tok/s |
| Gemma 2 27Bmid-size | Runs | Comfortable~11.0 tok/s |
| Gemma 3 27Bmid-size | Runs | Comfortable~11.6 tok/s |
| Qwen3.8-27Bmid-size | Runs | Comfortable~12.0 tok/s |
| Muse Glimmer 30Bmid-size | Runs | Comfortable~12.1 tok/s |
| NVIDIA Nemotron 3.5 Lightningmid-size | Runs | Not estimated |
| Qwen3 Coder 30B A3Bmid-size | Runs | Not estimated |
| Gemma 4 31Bmid-size | Runs | Comfortable~10.3 tok/s |
| Qwen3 32Bmid-size | Runs | Comfortable~9.3 tok/s |
| Qwen2.5 32Bmid-size | Runs | Comfortable~9.4 tok/s |
| Nanbeige4.3-33Bmid-size | Runs | Comfortable~9.1 tok/s |
| Qwen3.6-35B-A3Bmid-size | Runs | Not estimated |
| Recursive Flintmid-size | Runs | Not estimated |
| Llama 3.3 70Blarge | Runs | Slow~4.5 tok/s |
| DeepSeek-R1-Distill 70Blarge | Runs | Slow~4.5 tok/s |
| Qwen2.5 72Blarge | Runs | Slow~4.1 tok/s |
| NVIDIA Nemotron 3.5 Bolt 74B-A9Blarge | Runs | Not estimated |
| Kolibri-1large | Runs | Not estimated |
| Llama 4 Scoutvery large | Runs | Not estimated |
| gpt-oss-120bvery large | Runs | Not estimated |
| Laguna S 2.1very large | Doesn't fit | – |
| NVIDIA Nemotron 3.5 Super VLvery large | Runs | Not estimated |
| Ornith-1.5-122Bvery large | Runs | Not estimated |
| Qwen3.8-Flash-Nextvery large | Doesn't fit | – |
| MiniMax M3.1 Flashvery large | Doesn't fit | – |
| Reflection Beam 1.0very large | Doesn't fit | – |
| DeepSeek V4.1 Flash 1017very large | Doesn't fit | – |
| Meta Muse Spark Openvery large | Doesn't fit | – |
| Kimi K3.1 Provery large | Doesn't fit | – |
Not estimated: these models use only part of themselves for each word. Our speed estimate can't model that yet, so we leave it out instead of guessing.
Other quality levels
Higher quality needs more memory. Standard quality is what most people run at home.
- Q4 · standard (smallest that keeps quality)
- Q5 · a bit sharper, a bit bigger
- Q8 · near full quality, almost twice Q4
- FP16 · full quality, over three times Q4
Swipe the table sideways to see every quality level.
| AI model | Q4 | Q5 | Q8 | FP16 |
|---|---|---|---|---|
| LFM3-1.7B | Runs | Runs | Runs | Runs |
| Llama 3.2 3B | Runs | Runs | Runs | Runs |
| Spark-X2.6-4B | Runs | Runs | Runs | Runs |
| Mistral 7B v0.3 | Runs | Runs | Runs | Runs |
| NVIDIA Nemotron 3.5 Bolt E4B | Runs | Runs | Runs | Runs |
| Llama 3.1 8B | Runs | Runs | Runs | Runs |
| Gemma 2 9B | Runs | Runs | Runs | Runs |
| Qwen2.5 14B | Runs | Runs | Runs | Runs |
| Phi-4 14B | Runs | Runs | Runs | Runs |
| DeepSeek R1 Distill Qwen 14B | Runs | Runs | Runs | Runs |
| gpt-oss-20b | Runs | Runs | Runs | Runs |
| Mistral Small 3.2 24B | Runs | Runs | Runs | Runs |
| Gemma 4 26B A4B | Runs | Runs | Runs | Runs |
| OrcaCyber Nano 1.0 | Runs | Runs | Runs | Runs |
| Gemma 2 27B | Runs | Runs | Runs | Runs |
| Gemma 3 27B | Runs | Runs | Runs | Runs |
| Qwen3.8-27B | Runs | Runs | Runs | Runs |
| Muse Glimmer 30B | Runs | Runs | Runs | Runs |
| NVIDIA Nemotron 3.5 Lightning | Runs | Runs | Runs | Runs |
| Qwen3 Coder 30B A3B | Runs | Runs | Runs | Runs |
| Gemma 4 31B | Runs | Runs | Runs | Runs |
| Qwen3 32B | Runs | Runs | Runs | Runs |
| Qwen2.5 32B | Runs | Runs | Runs | Runs |
| Nanbeige4.3-33B | Runs | Runs | Runs | Runs |
| Qwen3.6-35B-A3B | Runs | Runs | Runs | Runs |
| Recursive Flint | Runs | Runs | Runs | Runs |
| Llama 3.3 70B | Runs | Runs | Runs | Doesn't fit |
| DeepSeek-R1-Distill 70B | Runs | Runs | Runs | Doesn't fit |
| Qwen2.5 72B | Runs | Runs | Runs | Doesn't fit |
| NVIDIA Nemotron 3.5 Bolt 74B-A9B | Runs | Runs | Runs | Doesn't fit |
| Kolibri-1 | Runs | Runs | Runs | Doesn't fit |
| Llama 4 Scout | Runs | Runs | Doesn't fit | Doesn't fit |
| gpt-oss-120b | Runs | Runs | Runs | Runs |
| Laguna S 2.1 | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| NVIDIA Nemotron 3.5 Super VL | Runs | Runs | Doesn't fit | Doesn't fit |
| Ornith-1.5-122B | Runs | Runs | Doesn't fit | Doesn't fit |
| Qwen3.8-Flash-Next | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| MiniMax M3.1 Flash | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| Reflection Beam 1.0 | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| DeepSeek V4.1 Flash 1017 | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| Meta Muse Spark Open | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| Kimi K3.1 Pro | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
Ran a model on this machine? Send us your llama-bench result.
Speeds and memory are our estimate from the hardware specs, not a measured test. Prices are dated and may change. How we calculate
Sources and method
Fit and speed are our estimate from memory size and memory speed, not a measured test. Checked . GEEKOM (Germany).
Price details and parts
New price at GEEKOM's web shop in Germany, in stock.
Source: GEEKOM (Germany), checked 21 Oct 2026.
€3,999 for the Ryzen AI Max+ 395 128 GB / 2 TB model at GEEKOM's German web shop on 21 Oct 2026, in stock (€4,299 struck through); we did not use a coupon price. GEEKOM lists 45–120 W, 154 W in its GEEK mode. AMD's Variable Graphics Memory (VGM) FAQ: up to 96 GB of a 128 GB Ryzen AI Max+ 395 for the graphics, the memory we count.
What this is
- GEEKOM A9 Mega, Ryzen AI Max+ 395, Radeon 8060S, 128 GB LPDDR5x, 2 TB SSD, as listed on GEEKOM's German web shop
Prices exclude shipping. Euro prices include VAT. Dollar prices are converted from our euro prices without EU VAT and before US sales tax, so they are an estimate.
Technical specs
- Shared memory
- 128 GB shared memory (96 GB of it for the graphics)
- Usable for AI models (our estimate)
- About 96 GB. The system keeps the rest for itself.
- Memory speed (bandwidth)
- 256 GB/s. How fast the chip can read the model; higher means faster answers. Source
- Power (maximum)
- 120 W. Source
- In short
- 128 GB of shared memory. AMD lets the graphics use up to 96 GB of it, and that is what we count.
Common questions
- Can the GEEKOM A9 Mega (Ryzen AI Max+ 395) 128GB run Llama 3.3 70B?
- It fits, but slowly: about 4.5 tokens per second (tok/s) by our estimate (Slow). A bit slower than you read. Fine for short answers. At standard quality and with a normal-length conversation, Llama 3.3 70B needs about 46 GB of memory, and the GEEKOM A9 Mega (Ryzen AI Max+ 395) 128GB has about 96 GB it can use.
- Which AI models run fully on the GEEKOM A9 Mega (Ryzen AI Max+ 395) 128GB?
- By our estimate, 35 of the 42 models we track run fully in its memory at standard quality: LFM3-1.7B, Llama 3.2 3B, Spark-X2.6-4B, Mistral 7B v0.3, NVIDIA Nemotron 3.5 Bolt E4B, Llama 3.1 8B, Gemma 2 9B, Qwen2.5 14B, Phi-4 14B, DeepSeek R1 Distill Qwen 14B, gpt-oss-20b, Mistral Small 3.2 24B, Gemma 4 26B A4B, OrcaCyber Nano 1.0, Gemma 2 27B, Gemma 3 27B, Qwen3.8-27B, Muse Glimmer 30B, NVIDIA Nemotron 3.5 Lightning, Qwen3 Coder 30B A3B, Gemma 4 31B, Qwen3 32B, Qwen2.5 32B, Nanbeige4.3-33B, Qwen3.6-35B-A3B, Recursive Flint, Llama 3.3 70B, DeepSeek-R1-Distill 70B, Qwen2.5 72B, NVIDIA Nemotron 3.5 Bolt 74B-A9B, Kolibri-1, Llama 4 Scout, gpt-oss-120b, NVIDIA Nemotron 3.5 Super VL, Ornith-1.5-122B.
- How fast is the GEEKOM A9 Mega (Ryzen AI Max+ 395) 128GB with AI models?
- With Qwen2.5 32B, the largest well-known model it runs well, about 9.4 tokens per second (tok/s) at standard quality (our estimate, not a measured test). Comfortable: About as fast as you read. Fine for chat.