Last week, 192GB mini PCs looked almost ridiculous.
GMKtec, Minisforum, Acemagic, Framework and others showed new systems around AMD’s Ryzen AI Max+ PRO 495. The processor supports 192GB unified memory, and AMD says as much as 160GB can be assigned to graphics. Suddenly a small desktop could hold models that would need several consumer GPUs in a normal PC.
Then DeepSeek released DeepSeek V4.1 Flash on September 10.
The new model has a 552B parameter backbone, activates only 8B parameters while reading input and 16B while generating output, supports a one million token context, and uses a new cache design that needs far less memory than the previous V4 Flash.
Sounds perfect for these new high memory mini PCs, right?
Then I checked the actual model files.
The official vLLM recipe puts the V4.1 Flash checkpoint at roughly 511GB on disk. Its recommended serving configuration budgets around 614GB of VRAM once headroom is included. The reference setups are not mini PCs. They are things like one GB200 NVL4 tray or an eight GPU H200 server.
So basically, the local AI hardware industry spent the first week of September celebrating 192GB. DeepSeek needed only a few days to make 192GB look small again.
Access without medium partner: 192GB Mini PCs Just Arrived
192GB was supposed to solve the memory problem
This is what made the Ryzen AI Max+ PRO 495 so easy to write about.
AMD supports a 256 bit LPDDR5X memory interface and up to 192GB of unified memory. GMKtec’s EVO-X5 Pro can assign up to 160GB to graphics. Minisforum announced the MS-S1 MAX-P495 and N5 MAX-P495 with the same 192GB ceiling. Acemagic put the chip inside a roughly 2 litre F9A mini workstation.
The number was big enough to change the normal GPU discussion.
An RTX 5090 has 32GB of GDDR7. A high memory AMD mini PC can give its integrated Radeon 8065S roughly five times that amount of usable graphics memory. Yes, the RTX 5090 is massively faster when the model fits. We already know that. The AMD box wins a different fight because it can keep much larger models in one memory pool.
AMD even says the 160GB graphics allocation is enough for 300B plus parameter models at 4 bit quantization.
And vendors were not only repeating a theoretical number. Acemagic says it demonstrated DeepSeek V4 Flash 284B locally on the new F9A PRO 495. That older V4 Flash model has 284B total parameters and only 13B active per token, so it was a very good example of why huge unified memory can matter even when the integrated GPU is slower than a high end NVIDIA card.
For about one week, 192GB felt like the new local AI sweet spot.
One week.
DeepSeek doubled the parameter count and made memory weird again
V4.1 Flash is not a normal step from 284B to something slightly larger.
DeepSeek lists the new model as a 552B parameter Mixture of Experts model. That number already sounds huge, but it does not tell the whole memory story.
The official vLLM recipe breaks the checkpoint into several parts. Routed experts plus the DSpark draft experts take about 259.5GiB. The new Engram tables take about 183.1GiB. Attention, dense projections, routers, embeddings, scales and other pieces add more.
Put it together and the checkpoint is roughly 511GB on disk.
I spent way too long checking this because Hugging Face also shows a larger parameter number in its file metadata, and for a minute I thought one of the sources was wrong. They are counting different things. DeepSeek’s 552B figure refers to the backbone, while the Engram memory tables, DSpark draft components and some other stored parameters sit outside that headline number.
This part is messy. AI model marketing usually gives us one parameter count and we pretend the model is one big clean block. V4.1 Flash makes that habit useless.
The machine does not care which parameters were included in the marketing number.
It has to store the files.
Only 8B active does not mean only 8B needs memory
This is probably the biggest misunderstanding people will have with V4.1 Flash.
DeepSeek says the model activates only 8B parameters per input token and 16B per output token. That is a big reason the model can be cheaper to serve than another model with hundreds of billions of active parameters.
But active parameters and resident parameters are not the same thing.
A Mixture of Experts model chooses a small set of experts for each token. It does not know in advance which experts every future token will need, so the wider set of experts still has to be accessible somewhere.
On a data center setup, that normally means keeping the model weights spread across a large pool of accelerator memory.
The official vLLM recipe budgets about 614GB of VRAM and lists verified H200, GB200 and GB300 configurations. One GB200 NVL4 tray with 768GB fits it. An eight H200 node gives 1,128GB. That leaves room for runtime state and batch work.
A 192GB mini PC is nowhere close to that official path.
And yet people will still run the model on consumer hardware. They already are.
That is where the story gets more fun.
Someone already ran it on one RTX 5090
Only days after V4.1 Flash landed, a community project published a llama.cpp port that runs the model using one RTX 5090 with about 31.8GiB of VRAM and roughly 125.7GiB of system RAM.
Read that again.
The official serving recipe wants hundreds of gigabytes of accelerator memory, yet a developer got it running with one consumer GPU.
The catch is storage and offload.
The project’s GGUF target is around 502GB across 11 shards. The system keeps some data resident and pulls other content through RAM and storage as needed. In the published results, new content ran around 5.12 tokens per second, while content already resident could reach about 21.27 tokens per second.
Those numbers come from one early community port, not a controlled comparison against the official server setup. Still, they tell us something useful.
“Can run” has become a nearly meaningless phrase unless the author also tells you where the model is stored, how much is resident, what gets offloaded, the quantization, context length, prompt speed and output speed.
A model can run on an RTX 5090.
It can also be a 500GB model that spends part of its life waiting on memory and storage.
Both statements are true.
This makes the 192GB mini PC more interesting, not less
At first I thought V4.1 Flash made the new AMD mini PCs look outdated before they even shipped. Well, sort of. The official model is far too large for the clean all-in-memory experience people imagined when they saw 192GB.
But a 192GB unified memory machine still has a useful advantage over the 5090 plus system RAM experiment.
The Radeon GPU can directly work from the same large memory pool used by the CPU. There is no fixed 32GB dedicated VRAM wall in the middle.
That does not mean a 192GB AMD box can simply load the entire 511GB checkpoint. It cannot.
It does mean a compressed community version might need much less offload than a 32GB discrete GPU setup. The exact result will depend on future quantization work and software support. We do not have a good independent PRO 495 V4.1 Flash benchmark yet, and I would ignore anyone giving precise tokens per second without showing the model files and settings.
The previous V4 Flash is a good clue. Community developers managed to fit the 284B model plus a speculative draft on a 128GB Ryzen AI Max+ 395 system and reported around 32 tokens per second in one custom ROCm setup. That required architecture specific work, aggressive compression and a custom inference engine. It was not a normal click Download in LM Studio kind of setup.
So yes, software can bend the memory rules quite far.
It cannot delete 511GB of weights.
DeepSeek fixed the KV cache and made the weights the bigger problem
There is a funny part to V4.1 Flash.
DeepSeek spent serious engineering effort reducing cache memory.
The new Causal Encoder Decoder architecture changes how the model handles the global KV cache. DeepSeek says V4.1 Flash needs only one quarter of the HBM and one eighth of the SSD storage for KV cache compared with V4 Flash.
The vLLM recipe says the compressed global KV works out to roughly 890 bytes per token. Even a full one million token prompt stays under 1GB for that global KV component, plus a fixed local sliding window per layer.
That is a wild reduction.
Then the checkpoint itself is about 511GB.
So the cache became much smaller, and now the weights are sitting there like a refrigerator in a studio apartment.
This is why architecture details matter. V4.1 Flash is built for giant serving systems where repeated agent prompts and long contexts can make cache cost painful. Cutting cache memory by this much can save real money at scale.
For a person staring at a 192GB mini PC, the bigger issue comes earlier.
The model itself does not fit.
192GB mini PCs are still a very good place for older giant models
I would not cancel a 192GB mini PC order because V4.1 Flash exists.
That would be like refusing to buy a 4TB SSD because somebody released a 10TB dataset.
The new machines still cover a huge part of the local model world.
DeepSeek V4 Flash 284B is one example. Large Qwen models, big coding MoE models, 70B dense models at high quality quantization, multimodal models, several smaller agents at once, large context workloads and local RAG systems can all benefit from having one large memory pool.
And 192GB does not have to be used as 160GB of graphics memory all the time.
A local agent server needs room for the model plus the operating system, vector database, browser, code tools, embeddings, document processing and maybe a second smaller model doing routing or vision work. Giving the whole machine 192GB is useful even if the main LLM only uses half of it.
This is where I think the mini PC marketing should focus.
“Runs 300B” gets clicks.
“Can keep a serious local AI setup alive without memory juggling all day” is the reason somebody may keep using the machine after the demo ends.
The model race is moving faster than the hardware memory race
Consumer GPU memory moves painfully slowly.
The RTX 3090 launched with 24GB in 2020. The RTX 4090 still had 24GB in 2022. The RTX 5090 finally moved to 32GB.
Six years. Eight extra gigabytes.
On the AMD unified memory side, things have moved much faster. Ryzen AI Max+ 395 systems reached 128GB. The new PRO 495 supports 192GB. Vendors are already advertising 160GB for graphics.
Then model developers come along and casually add hundreds of gigabytes.
V4 Flash was 284B. V4.1 Flash has a 552B backbone plus a giant Engram memory component. The official checkpoint is roughly half a terabyte.
Hardware buyers are chasing a moving target.
I think this is why the obsession with “future proofing” a local AI machine is kind of hopeless. A 192GB box looks absurd today and small the next week, depending on which model you pick.
Buy for the models you plan to use.
That sounds boring. It also saves money.
DeepSeek V4.1 Flash itself is stranger than the size headline
The model is not big just because DeepSeek wanted a bigger number.
Its new architecture is built around asymmetric work. Input activates about 8B parameters. Output activates about 16B. DeepSeek also added Engram memory, a large conditional n-gram memory system that stores patterns outside the normal backbone.
This is part of why Hugging Face and the serving recipe can show numbers that look confusing beside the 552B headline.
The model also handles images and text natively and keeps the one million token context length.
DeepSeek says V4.1 Flash beats the older V4 Pro on its own benchmark set despite being the smallest member of the new architecture family. The company has already retired the older V4 Flash endpoints, and from September 14 it is routing V4 Pro API requests to V4.1 Flash until V4.1 Pro arrives.
That last part is happening right now.
The model is four days old as I write this. Local runtimes are still catching up. Quantizers are already repacking parts of it. The llama.cpp community has a working port. vLLM has official NVIDIA and AMD recipes aimed at data center hardware.
In another month, the local story may look different again.
The next 192GB benchmark I want is obvious
Someone needs to put V4.1 Flash on a Ryzen AI Max+ PRO 495 system and stop talking in parameter counts.
I want to see the full setup.
Which quantization fits? How much of the 511GB checkpoint stays in unified memory? How much sits in system storage? What is the prompt processing speed? What happens on fresh content versus resident content? Does 160GB assigned to graphics help, or does a more flexible shared allocation work better? How does ROCm compare with a llama.cpp path?
And please show power draw.
A small 120W class APU system pulling a half terabyte model through memory and storage is a very different thing from eight H200 GPUs. Nobody expects the same throughput. The useful comparison is how much capability you can get from one desk without a rack.
Right now we do not have enough clean numbers.
That is the frustrating part. The hardware announcements are ready. The model is ready. The benchmark table I actually want does not exist yet.
511GB is the new reminder that “local” has levels
Local AI used to mean running a 7B model on a laptop.
Then it meant a 70B model on a workstation.
This year we started calling 284B MoE models local because clever quantization and unified memory made them usable on one small computer.
Now V4.1 Flash is a roughly 511GB checkpoint.
Somebody has already pushed it onto one RTX 5090 with RAM and storage offload. The official deployment path wants more than 600GB of VRAM. A 192GB mini PC sits somewhere in the middle.
All of those count as local.
They do not give the same experience.
That is the part hardware articles need to get better at explaining. A model fitting on local storage is one level. A model fitting in system memory is another. A model sitting fully inside accelerator accessible memory is another. Then there is the question of whether it runs at 2 tokens per second or 30.
“Local” tells you where the model is.
It does not tell you whether you will enjoy using it.
I still think 192GB is the right direction
After all this, I am more interested in the new AMD machines, not less.
The RTX 5090 approach gives local AI users huge speed and a small memory ceiling. The 192GB mini PC approach gives us a much larger ceiling and lower bandwidth. Both have obvious problems once models start reaching half a terabyte.
But memory capacity is the problem consumer hardware has ignored for too long.
DeepSeek V4.1 Flash makes that painfully clear.
A week ago, 160GB of graphics memory in a mini PC sounded excessive. Today, the official V4.1 Flash checkpoint is more than three times that size.
Maybe nobody should try to run every new frontier model on a desk. Fair enough.
But open weights keep getting better, and people will keep trying anyway.
That is why this hardware race is getting so fun to watch. Every time the PC industry gives local AI more memory, model builders find a way to use all of it and ask for more.
The 192GB mini PC is not old.
DeepSeek just gave it a very big next problem.