128GB for Local AI: Upgrade Your Gaming PC or Buy a Workstation?

128GB has become the magic number in local AI. It is the amount people mention when they want to run a huge model at full context without calling an API. It is also the point where the shopping list gets silly very quickly.

I watched Kai compare every practical route to 128GB, from Strix Halo mini PCs and a Mac Studio to a DGX Spark and a four-GPU server. The part that changed my mind was not one of the expensive machines. It was a gaming PC with a 12GB graphics card and 64GB of ordinary RAM.

That PC cannot turn 12GB of VRAM into 128GB. What it can do is split the work intelligently. With Qwen3.8 Flash-Next and Strata, the active part stays close to the GPU while the rest sits in system memory. For the right model and quantization, that is enough to reach genuinely useful speed.

My short answer

Your situation	What I would do	Why
You already own a PC with a 12GB GPU	Upgrade to 64GB of RAM and test Strata	The 3-bit builds reached 50 to 62 tokens per second in Kai's RTX 5070 example
You want 4-bit quality in one compact machine	Look at a 128GB Strix Halo system running Linux	It was the least expensive complete machine in the video that ran the 4-bit build well
You need MLX, CUDA or a polished workstation	Choose the software stack first, then compare Mac Studio and DGX Spark	The premium only makes sense when the ecosystem is part of the requirement
You use a large model only a few hours each week	Rent the compute	An expensive local box will spend most of its life idle

If a capable gaming PC is already under your desk, try the RAM upgrade before shopping for a $3,000 workstation.

Why 128GB became the target

The model at the center of Kai's test is Qwen3.8 Flash-Next. It has 125 billion model parameters, plus a 51 billion parameter lookup table for common token sequences. The 4-bit files shown in the video sit between roughly 94GB and 124GB before adding the memory required for context and the inference runtime.

That is why 128GB sounds ideal. The headline capacity looks large enough for the weights while leaving one unified pool for the CPU and GPU. The problem is that installed memory and memory available to the model are not the same number.

The 128GB label hides several different limits

Unified memory is shared. The operating system needs some, the desktop needs some, and the GPU runtime may impose its own ceiling. Kai's examples make the mismatch easy to see.

Platform shown in the video	Installed memory	Memory available to the workload	Important detail
Ryzen AI Max on Windows 11	128GB	Up to 96GB by default	The GPU cannot automatically claim the entire pool
Ryzen AI Max on Linux	128GB	Close to 120GB in owner reports cited by Kai	Kernel, driver and firmware configuration matter
Apple Silicon	128GB	Around 96GB at the default limit	The allocation limit can be raised, but the system still needs memory
NVIDIA DGX Spark	128GB	About 122GB	Kai notes that the 124GB file still needs offloading
Windows RTX Spark systems	128GB	Not stated in the material Kai found	I would wait for independent model-specific testing

I would never buy one of these machines from the memory figure alone. I would first find a test using the same operating system, runtime, model file and context length that I plan to use. A 128GB machine with a 96GB allocation limit can fail exactly where a buyer expected it to shine.

Why a 12GB GPU can still be useful

Qwen3.8 Flash-Next is sparse. The full model is huge, but only about 6 billion parameters are active for each generated token. The rest still has to remain accessible, but every byte does not need the fastest path for every token.

Strata is built around that behavior. It keeps the latency-sensitive work on the GPU and uses system RAM for the larger set of inactive experts and the lookup table. Kai compares it to a small desk beside a bookshelf: the desk holds the pages being used now, while the shelf keeps the rest within reach.

This is not a universal trick for every large model. The result depends on Flash-Next's architecture and on an engine designed to exploit it. Point the same PC at a dense model or a different sparse design and the performance can change completely.

The RTX 5070 result is the important one

Kai highlights the reference system from the Strata repository: an RTX 5070 with 12GB of VRAM, a Ryzen 5 7600 and 64GB of RAM. Nothing in that list resembles a 128GB AI workstation, yet the measured speeds are competitive with machines costing several thousand dollars.

Build	Generation	Prompt processing
Q2_0	94 tokens/s	2,650 tokens/s
IQ2_XS	79 tokens/s	2,090 tokens/s
IQ3_XXS	62 tokens/s	1,750 tokens/s
IQ3_S	53 tokens/s	1,620 tokens/s
Coder	55 tokens/s	2,180 tokens/s

The 3-bit results are the sweet spot here. They are fast enough for an interactive coding or research session, and prompt processing at 1,620 to 1,750 tokens per second is especially useful when a tool keeps rereading a large codebase.

The compromise is precision. Kai cites a lab result claiming that a 3-bit file retains 99.4% of the full model's score across three benchmarks, but that does not settle the question for long coding sessions. There is still no convincing side-by-side test covering the exact work that would make someone buy a 128GB machine. I would test my own prompts before paying thousands of dollars for the fourth bit.

What the expensive routes actually buy

The prices below are the figures quoted in Kai's video. They are a snapshot, not live store prices. Availability and software support can move faster than hardware reviews.

Route	Price quoted by Kai	Speed quoted in the video	What I think it is for
Existing 12GB gaming GPU with 64GB RAM	The cost of a RAM upgrade	50 to 62 tokens/s at 3-bit	The first option I would test for one user
128GB Strix Halo mini PC	From $3,099	47 to 54 tokens/s at 4-bit with a suitable engine	A compact 4-bit machine, preferably on Linux
128GB Mac Studio	$5,099	About 40 tokens/s for prose and 75 for code in the cited MLX test	A quiet MLX workstation when the Apple workflow matters
DGX Spark	$6,950	About 41 tokens/s median and 27 on prose in Kai's comparison	A CUDA purchase, not the value choice for one model
Four used Radeon Pro V620 cards	About $3,000 for the complete build at the time	60 to 70 tokens/s at 4-bit	A noisy, power-hungry server project or a multi-user box

The engine can matter more than the logo on the case. In Kai's Strix Halo example, one engine processed a 160,000-token prompt at 113 tokens per second while another reached 1,227 on the same machine and model. That is roughly 23 minutes versus 2 minutes just to read the prompt.

The multi-GPU server wins on raw throughput, but it also pulled 500 to 600 watts while generating. It made sense when the used cards were cheap. Once those cards became popular and prices rose, the bargain became much harder to reproduce.

What I would buy

I already have a gaming PC

I would check Strata's supported hardware, install 64GB of RAM, leave enough SSD space for the model files and test a 3-bit build. I would record generation speed, prompt ingestion speed, context limit and answer quality on my own work. That experiment is cheap compared with replacing the whole machine.

I want one compact 4-bit box

I would choose a 128GB Strix Halo system and run Linux. I would also decide on the inference engine before choosing the brand of mini PC. The video shows too large a software gap to treat the runtime as an afterthought.

I need Mac or CUDA software

I would buy the Mac Studio for an MLX workflow, quiet operation and the rest of the Apple ecosystem. I would buy DGX Spark only when a real part of my workload requires CUDA. Neither would be my default recommendation based on tokens per dollar.

I run several agents or serve several people

This is where a larger workstation or multi-GPU server starts to make sense. A single 12GB gaming card is an excellent one-person setup, but aggregate throughput and concurrent streams change the calculation.

I only use local AI occasionally

I would rent. Kai compares a $3,099 Strix Halo with a 96GB cloud GPU at $2.09 per hour. The purchase price covers about 1,483 rental hours before electricity, maintenance or depreciation. If the machine will be idle most of the week, owning it is mostly paying for convenience and privacy.

The conclusion I took from Kai's test

128GB is useful, but it is not one product and it is not one performance level. A 128GB label can mean 96GB available to the GPU, 122GB available, or a pool whose speed depends heavily on the engine. The model architecture matters just as much as the capacity.

For Qwen3.8 Flash-Next, the most interesting setup is the ordinary one: a supported 12GB GPU, 64GB of RAM and Strata. It trades the fourth bit for a much lower entry price and still produces 50 to 62 tokens per second in the cited 3-bit tests.

If that trade is not good enough for my workload, I would move to Strix Halo. I would not jump directly to a $5,000 or $7,000 machine unless its software ecosystem solves a problem I can name.

---

Sources

• Kai: Every Ways to Get 128GB VRAM for Local AI at Full Context
• Strata repository and benchmark notes
• Qwen3.8 Flash-Next model card
• Qwen3.8-Next architecture paper
• NVIDIA DGX Spark specifications