
Pictured: Corsair 2000D RGB AIRFLOW (black). The Framework Desktop mainboard inside is not shown.
CG AI Halo 128 Workstation
Black Friday price: you save $306.77 until Dec 1. No code needed.
Paid in Bitcoin, Monero, USDT or another coin you choose.
Free insured shipping. US delivery in about 5 to 11 business days. Shipping details
Overview
CG AI Halo 128 is a Strix Halo build centered on Framework's standalone Desktop Mainboard: a Ryzen AI Max+ 395 with 128GB of soldered LPDDR5X. We fit it in a Corsair 2000D RGB Airflow Mini-ITX case with a Corsair SF750 SFX power supply, a Noctua 120mm fan on the board's heatsink and a 2TB WD_BLACK SN850X.
The CPU and GPU share all 128GB at about 256GB/s, and the graphics side can address up to 96GB of it. A 70B model at 4-bit uses roughly 40GB and fits with long context, but dense models that size generate only a few tokens per second at this bandwidth. Mixture-of-experts models around 120B, such as gpt-oss-120B, are the better match.
It suits a low-power inference box for one or two users, not training or CUDA-only software. The memory cannot be upgraded later, and the mainboard offers a single PCIe x4 slot, so a discrete graphics card would be bandwidth-limited. The tower is much larger than Framework's own 4.5L chassis.
- Ryzen AI Max+ 395, 128GB LPDDR5X
- About 256GB/s memory bandwidth
- 2TB NVMe in a Mini-ITX tower
- 70B at 4-bit, 120B MoE
Specifications
| Processor | AMD Ryzen AI Max+ 395, 16 Zen 5 cores / 32 threads, up to 5.1GHz boost, 64MB L3 |
|---|---|
| Graphics | Integrated Radeon 8060S, 40 compute units |
| NPU | XDNA 2, up to 50 TOPS |
| Memory | 128GB LPDDR5x-8000, 256-bit, soldered to the mainboard |
| Memory bandwidth | About 256GB/s |
| GPU memory | Up to 96GB addressable by the GPU |
| Storage | WD_BLACK SN850X 2TB, M.2 2280 PCIe 4.0 x4, up to 7,300MB/s read |
| Local AI capacity | 70B at 4-bit with long context; 100B to 120B MoE at 4-bit |
| Mainboard | Framework Desktop Mainboard, Mini-ITX, Ryzen AI Max+ 395 with 128GB |
|---|---|
| Cooling | Board heatsink with Noctua NF-A12x25 PWM 120mm fan |
| Power supply | Corsair SF750 (2024), 750W SFX, ATX 3.1, Cybenetics Platinum, 7-year warranty |
| Expansion | 1x PCIe x4 slot on the mainboard |
| Case | Corsair 2000D RGB AIRFLOW, Mini-ITX, 458 x 271 x 200mm, 165mm cooler clearance |
|---|---|
| Case fans | 3x AF120 RGB SLIM included |
| Front I/O | 1x USB 3.2 Gen 2 Type-C, 2x USB 3.2 Gen 1 Type-A |
| Ethernet | 5Gbit (Realtek RTL8126) |
| Mainboard ports | 1x HDMI 2.1, 2x DisplayPort 2.1, 2x USB4 Type-C, 2x USB-A 3.2 Gen 2, 3.5mm combo jack |
| Operating system | Ubuntu 24.04 with Ollama, llama.cpp and Open WebUI, Ubuntu 24.04 with open models preloaded, or Windows 11 Pro (+$189.05) |
|---|---|
| Warranty | 3-year parts, lifetime support |
Questions and answers
Still unsure? Visit the help center or message us.
Which case and power supply does this build use?
A Corsair 2000D RGB Airflow Mini-ITX case, which has 165mm of cooler clearance and takes SFX power supplies, with a Corsair SF750 (2024). The Framework mainboard follows Mini-ITX conventions, so it mounts like a standard board.
Which model sizes fit in the 128GB pool?
A 70B model at 4-bit needs about 40GB and runs with long context. Mixture-of-experts models around 120B fit at 4-bit and generate much faster than dense models of the same size.
Why is dense 70B generation slow?
Each generated token reads the whole set of weights, and 256GB/s of bandwidth limits that to single-digit tokens per second for a 40GB model. The discrete cards in our other builds offer 456GB/s to 1,792GB/s per card, but each holds far less than 128GB.
Can I expand it later?
The 128GB of LPDDR5X is soldered and fixed. The mainboard has two M.2 slots and one PCIe x4 slot, so storage is easy to add, while a graphics card would run at x4 bandwidth.
How many models can the 2TB drive hold?
A 4-bit 70B model is about 40GB, so the drive holds dozens of them alongside the operating system. WD rates the SN850X at up to 7,300MB/s sequential read, which keeps load times short.
From the blog

How much VRAM for local LLM work in 2026: a practical sizing guide
A practical 2026 sizing guide: add up weights, KV cache and overhead, check real GGUF sizes for current open models, and see what fits on 16GB to 512GB.

Best computer for local AI: DGX Spark vs Mac Studio vs Strix Halo
NVIDIA DGX Spark, Mac Studio M5 Ultra and Ryzen AI Max+ 395 compared on memory, bandwidth, real benchmarks and price, plus when a GPU tower is the better buy.
Pairs well with
Picked to go with this product. Check sizes and connectors for your build before you order.







