What MSI says the machine is
MSI describes the EdgeMesa N AI+ Series as a compact AI desktop built on NVIDIA RTX Spark. It claims up to 1 PFLOP of FP4 AI performance and 128GB of unified memory. The RTX Spark N1X pairs a 6,144-core Blackwell RTX GPU with a 20-core Grace CPU. MSI's US store lists the CPU as 20-core Arm, with 10 Cortex-X925 and 10 Cortex-A725 cores. It lists the 128GB model's memory bandwidth as 273 GB/s.
The cooling is a vapor chamber with dual fans. MSI says this helps the system sustain 140W TDP performance and supports reliable 24/7 AI agent operation. MSI's store page also gives noise figures of 28 at idle and 32 at full load, with no unit stated. The reviewer's own description is simpler: he says he never heard the fans kick on.
Reviewer and vendor differ on a few specs:
- Volume: MSI's marketing says 1.3L. The review's spec table says about 1.2L and notes MSI lists 1.3L.
- Bandwidth: The review says roughly 300GB/s. MSI's store listing says 273 GB/s, which is the better figure to use.
- CPU clock: The review says up to about 4GHz. MSI's store listing says 3.25GHz.
Configurations and price
Micro Center lists a 128GB, 4TB model with Windows 11 Pro, 10GbE, Wi-Fi 7 and Bluetooth 5.4. That page shows a price of $7,499.99. The review reports two configurations:
| Configuration | Review-reported price | Notes |
|---|---|---|
| 64GB / 1TB | $4,399 | A WindowsForum report says MSI's US store showed this model at $4,399 and in stock. |
| 128GB / 4TB | $7,499 | The reviewed unit. |
The review says both prices were cut this week from $4,699 and $7,999. I couldn't confirm that cut independently. It names B&H Photo, Micro Center and MSI direct as sellers. Prices on a launch-week product can move, so check before you buy.
What the reviewer actually did with it
He moved all of his local AI work to the box and runs it headless, with remote desktop enabled as a fallback.
- Two models at once: He runs Qwen3.8-27B as the main agent and Qwen3.6 35B-A3B as a faster, lighter model. The lighter model handles Home Assistant integration and simple chats.
- Remote access: He exposes the main model on his network and over Tailscale through OpenCode. That lets him use it from apps like T3 on an iPad.
- Agent work: He uses Hermes as his agent, with the models served through LM Studio. He says he's still tuning that setup.
- Video editing: The EdgeMesa pulls footage from his NAS, logs it, runs HyperFrames and does the edit with ffmpeg. Claude Opus 5.5 drives the edit, so this is not an offline workflow. A 15-minute 1080p video renders in about seven minutes, similar to his Strix Halo machine. The benefit he stresses is that his main desktop stays responsive.
This is a personal deployment, not a vendor-supported recipe. Exposing a model endpoint beyond your LAN deserves its own security review.
Memory is the point
The reviewer says the GPU can address up to 105GB of the system's RAM. He fully offloaded an 88GB Qwen3.8-flash-next model to the GPU. His earlier test unit was an HP OmniBook Ultra 16 with 32GB of RAM, which he says forced constant compromises.
Speeds are usable but not blazing. He saw about 30 to 50 tokens per second on larger models like Qwen3.8-27B. In real Hermes calls through LM Studio he saw about 23 tokens per second. With both models loaded, 40 to 60GB of RAM stayed free for other work.
He is clear that this is not an RTX 5090 replacement. The 5090 has 32GB of GDDR7 and about 1.8TB/s of bandwidth. The EdgeMesa trades speed for capacity. The review doesn't state quantization, context length or runtime versions. Treat those token rates as one user's observations.
CUDA versus Strix Halo
Using the review's numbers, the Spark on CUDA compares to AMD's Ryzen AI Max+ 395 on its best backend like this:
| Workload | RTX Spark (CUDA) | Strix Halo |
|---|---|---|
| Qwen3.5 35B-A3B | 65.7 tok/s | 62.3 tok/s |
| gpt-oss-120b | 56.3 tok/s | 55 tok/s |
| Llama 3.3 70B | 5.73 tok/s | 4.54 tok/s |
| SDXL 1024px | 21.9 s/image | 37.7 s/image |
On the shared Vulkan runtime, he says Strix Halo wins clearly. His argument is that this isn't a fair fight, since the Spark is built around CUDA.
Text generation is close, which fits the similar memory bandwidth on both platforms. The real gap shows up in prompt processing and image generation. If you mostly generate text, the Spark's edge is small. If you do image work, it's substantial. These are the reviewer's own tests, with no settings published for them.
The Windows on Arm catch
The review's main caveat is software. Strix Halo is x86, so everything runs on it. The N1X is Arm, so some game anti-cheat systems won't run. Some tools, such as vLLM, are less mature on Arm. He says this is the only reason the EdgeMesa isn't his main desktop. He relies on too many apps that aren't native on Arm, and he doesn't want to deal with x86 emulation.
Gaming isn't the target use, but he tested one title. Forza Horizon 6 ran at 1440p on the Extreme preset at over 100 FPS with frame generation and over 60 FPS without it.
On Linux, he says there is no official option. He also says he never felt the need to install it during the week. That is his experience on this machine, not a claim about every Linux install.
Ports and clustering
I/O is thin: four USB-C 20Gbps ports (one used for power), one HDMI 2.1, and one 10GbE port. The platform spec sheet and MSI's listing both mention Wi-Fi 7 and Bluetooth 5.4. Three of the USB-C ports support DisplayPort Alt Mode 1.4, and one is for power input. There is no USB4, no Thunderbolt and no second NIC. He also notes there is no 200G ConnectX link like the one on NVIDIA's DGX Spark. That makes clustering several units impractical.
Power comes from a 240W USB-C adapter. He used only the official one. His guess that another 240W charger would work is an assumption, not tested guidance.
Who should buy it
Good fit:
- You want to host larger local models and agents at home with no token limits.
- You want a small, quiet, always-on Windows box.
- Your workloads use CUDA, especially image generation.
Skip it if:
- You need x86-only apps.
- You need Linux.
- You want to cluster multiple units.
- You want a daily-driver desktop.
The reviewer also says the models it runs sit in the "Sonnet" class, not the top frontier tier. That's a subjective comparison. At $7,499, the real question is whether your workflow runs on Arm. Check that first.
References
- MSI EdgeMesa N AI+ review: 128GB of RTX Spark makes a quiet home AI server How-To Geek · 2026-11-19T19:30:15+00:00
- NVIDIA Home nvidia.com
- MSI EdgeMesa N AI+ Series with NVIDIA RTX Spark msi.com