This is more of an ad, not a review, and reads like the author has hardly any experience with the things he's trying out. That Z Image Turbo diffusion model would've also run on many consumer GPUs and with way higher performance for a fraction of the price. Misleading.
Highly recommend lemonade server if you have a strix halo desktop. Been using Qwen3.6-35B @ Q_8 as my main driver and it’s been great with 60 TPS for generation. I occasionally use the 27B @ q6 but only get 20-25 TPS for generation with MTP.
I used llama swap for a while before leaning into lemonade. The UI has improved a lot, but be careful as the most of the models default to very small 4K context windows by default.
They’re doing some nice things with their Halo models, which load an ensemble of different types of model at the same time. With high vram it’s easy to keep them all in memory, so even though the compute is limited the context switching is fast.
You do lag a bit on the upstream engine releases, the llama.cpp/sd.cpp/whisper libraries are downloaded from inside the app.
vLLM is in experimental mode, I haven’t tested it. It’s limited in the models they suggest, but you can download anything from huggingface with a two click install.
Is there some means to monitor the queries it's sending (or hold and review before transmission) or throttle to avoid triggering abuse thresholds on any single domain?
Lemonade proxies your request to llama.cpp that it installs and manages. Is out of the box all local LLMs, but you can connect it to an API endpoint (supports OpenAI compatible or Anthropic compatible).
I think it stores basic logs but if you wanted to monitor you’re probably going to want to proxy it with an LLM gateway.
It seems like there is a very health space for an MOE targeted GPU where it has essentially an 5070ti with 16gb ish GDDR7 but then also has 128 GB LPDDR5x (or even just DDR5 as expansion dimms on it?). Putting this into the same card would likely reduce the transfer hit when a cache miss happened and the gpu had to load from the slower LPDDR5x. No need to have PCIe 5x16 limiting memory transfer if it is on the card. MOE models could then get near native performance and even models where the active parameters + context didn't quite fit the thrashing would be less of a problem. Not UMA but gets the UMA 'lots of system memory to play with' benefit.
Unfortunately, the table of models and tokens per second (TPS) and time to first token (TTFT) is not helpful without specifying the quantization of the model.
Does anyone else feel like it would be great to be able to purchase $4000 AI box but 128 gigs is not enough. If I spend all that money and it doesn’t really do what I wanted to do, whats the point?
It’s kind of like general aviation where you can go buy a Cessna but it’s only going to realistically get you somewhere you could drive anyways but do you really wanna spend that mush cash to get road trip distance at slightly better than road trip speeds? You really need a 5 million dollar jet and that’s just not practical. That’s sort of how I feel about this device.
I've got a Framework with 128GB. Sure, I'm probably not going to run Deepseek V4 flash (though, I could run the 2-bit quant). But there are a lot of models (especially MOE models) that run fine. Even Qwen3.5-122B in a 4 or 5 bit quant can run. The Qwen3-coder-next (~80B IIRC) runs fine (IIRC I'm running a 6-bit quant of that one). The much vaunted Qwen3.6-27B is runnable, but kind of slow (~20t/s with MTP), though that's not a memory limitation issue.
Yeah, It's be great to have 256GB RAM, but that's really expensive now anyway and there's no way I could get my wife to sign off on spending 5 grand (or more now) on a box with that much memory.
I own one, I don’t feel like the RAM is a huge issue (of course I want 192GB to run something like DS4 Flash). The lack of FP4 and slow memory bandwidth is rough. NVFP4 support is such a huge advantage that I would recommend others to buy a DGX spark over a strix halo if you’re using it purely for AI. Strix halo works better for general computing.
How has it been and how’s the speed? I read online that the it’s ~200 TPS for PP and ~15 TPS for TG. Unfortunately for those speeds, it’s very very hard to use for agentic stuff.
That's pretty accurate to what I've seen, I'd definitely recommend a smaller model for active agentic use on that hardware.
It definitely seems to be the leader on 'general intelligence' on this hardware from my casual usage, but the newer Qwen or Gemma series models are much more usable speed-wise for agentic use, and often is as good or better than DS4 on that front.
I saw that but end of the day, the chips themselves don’t have hardware support for FP4. There’s smart ways around this limitation but it will never natively be close to true FP4 performance like MXFP4 and NVFP4 (happy to be proven wrong though).
The Strix Halo is a great dev machine and a mediocre AI machine. You can run Qwen 3.6 27B at a decent speed, or larger MoE models, and that's about it. For some that's more than enough though, myself included.
Until RAM prices drop and can economically get machines with 256GB, 512GB and higher bandwidth... I frankly think the local AI story is going to be still fairly muted for most people.
My Spark can do Qwen3.6 MoE A3B at 60 to 70-ish token/second and that's really good, but there's limits the usefulness of that model. It's not useful for coding, in any case.
Once people can run something like GLM 5.2 at lower quants (512GB could do a passable job), then I think the story changes.
Whether we ever see DRAM as cheap as it was ever again, I don't know.
Agree - the 128GB Strix Halo is capable if you use LLMs as assistants, but it's not so good if you use LLMs as agents (or worse, agent teams/swarms) since all of the models that can fit on it are pretty dumb compared to frontier or near-frontier models. You can at best hope for Sonnet-level capabilities.
That doesn't mean that local models are useless though! If Mythos/Sol is an ASI that threatens to take your job and turn you into paperclips, then Qwen/Gemma is an old-fashioned office secretary that loyally helps you with tasks but doesn't have a good grasp of details. Every white-collar worker 50 years ago would have killed to have a hard-working personal secretary.
Exactly this. I own the Framework desktop board. I knew all of its limitations before I bought it, and it's ok to play with on a hobby level, but it isn't much more than a Radio Shack toy.
That memory bandwidth is painful. It's like trying to fill an Olympic swimming pool with a thimble.
Part of me wonders, would 3d xpoint (if still around) be a viable option?
Yeah it is slower than real RAM by a good amount for latency, but you can get similar bandwidth and the cost was history about half of the same size DDR.
Possibly - I've heard anecdotal reports that old Intel Optane chips are in hot demand right now. Intel/Micron probably would have made a killing if they had kept that product line alive for a few more years. Never miss an attempt to snatch defeat from the jaws of victory!
I was thought experimenting the other day... ~10 nVME drives striped and running parallel could approximate the memory bandwidth of DDR5 DRAM in a box like this. Like you say, latency wouldn't compete but on raw throughput would be comparable.
Not anymore cost effective, I guess, but gets you the ability to work over very large model sizes maybe. But the problem is that tensor matmul etc hardware wouldn't work effectively with it.
We're maybe only 2 years away from really useful, relatively affordable local LLM usage. You can buy a 5090 PC for $5-6k but 32GB of VRAM really limits model sizes to ~31B. And that won't change (even with NVidia's next generation) because NVidia uses VARM as an aggressive market segmentation technique.
No, the hope really is these other platforms with a shared memory architecture. The DGX Spark won't be it because of the aforementioned market segmentation. So that leaves two players: AMD and Apple.
The AMD platform is still too low memory bandwidth, currently <300GB/s. For comparison a 5090 (or 6000 Pro) is 1.8TB/s and the M3 Ultra Mac Studio is ~900GB/s. Oh and B100/B200 uses HBM3e memory at ~3.2TB/s. The M5 Max in some Macbook Pros tops out at ~600GB/s. So you need access to better RAM and better CPU architecture for all this.
My great white hope is Apple. They have the market power to get memory and build silicon that coul dhave enough FLOPS to compete with NVidia's platform. They've started talking about it and I've seen rumors they're targeting the M7 generation (2028) for a huge leap. I'll believe it when I see it however.
But the point is, I think we'll be running 31B models at 100+tok/s on enthusiast hardware in 2 years and we'll likely be able to locally run 100-400B models, possibly larger.
This is more of an ad, not a review, and reads like the author has hardly any experience with the things he's trying out. That Z Image Turbo diffusion model would've also run on many consumer GPUs and with way higher performance for a fraction of the price. Misleading.
It's Microcenter. Don't get your reviews from a retailer.
Highly recommend lemonade server if you have a strix halo desktop. Been using Qwen3.6-35B @ Q_8 as my main driver and it’s been great with 60 TPS for generation. I occasionally use the 27B @ q6 but only get 20-25 TPS for generation with MTP.
I second this.
I used llama swap for a while before leaning into lemonade. The UI has improved a lot, but be careful as the most of the models default to very small 4K context windows by default.
They’re doing some nice things with their Halo models, which load an ensemble of different types of model at the same time. With high vram it’s easy to keep them all in memory, so even though the compute is limited the context switching is fast.
You do lag a bit on the upstream engine releases, the llama.cpp/sd.cpp/whisper libraries are downloaded from inside the app.
vLLM is in experimental mode, I haven’t tested it. It’s limited in the models they suggest, but you can download anything from huggingface with a two click install.
Do you give it access to the internet?
Is there some means to monitor the queries it's sending (or hold and review before transmission) or throttle to avoid triggering abuse thresholds on any single domain?
Lemonade proxies your request to llama.cpp that it installs and manages. Is out of the box all local LLMs, but you can connect it to an API endpoint (supports OpenAI compatible or Anthropic compatible).
I think it stores basic logs but if you wanted to monitor you’re probably going to want to proxy it with an LLM gateway.
It seems like there is a very health space for an MOE targeted GPU where it has essentially an 5070ti with 16gb ish GDDR7 but then also has 128 GB LPDDR5x (or even just DDR5 as expansion dimms on it?). Putting this into the same card would likely reduce the transfer hit when a cache miss happened and the gpu had to load from the slower LPDDR5x. No need to have PCIe 5x16 limiting memory transfer if it is on the card. MOE models could then get near native performance and even models where the active parameters + context didn't quite fit the thrashing would be less of a problem. Not UMA but gets the UMA 'lots of system memory to play with' benefit.
Unfortunately, the table of models and tokens per second (TPS) and time to first token (TTFT) is not helpful without specifying the quantization of the model.
Does anyone else feel like it would be great to be able to purchase $4000 AI box but 128 gigs is not enough. If I spend all that money and it doesn’t really do what I wanted to do, whats the point?
It’s kind of like general aviation where you can go buy a Cessna but it’s only going to realistically get you somewhere you could drive anyways but do you really wanna spend that mush cash to get road trip distance at slightly better than road trip speeds? You really need a 5 million dollar jet and that’s just not practical. That’s sort of how I feel about this device.
I've got a Framework with 128GB. Sure, I'm probably not going to run Deepseek V4 flash (though, I could run the 2-bit quant). But there are a lot of models (especially MOE models) that run fine. Even Qwen3.5-122B in a 4 or 5 bit quant can run. The Qwen3-coder-next (~80B IIRC) runs fine (IIRC I'm running a 6-bit quant of that one). The much vaunted Qwen3.6-27B is runnable, but kind of slow (~20t/s with MTP), though that's not a memory limitation issue.
Yeah, It's be great to have 256GB RAM, but that's really expensive now anyway and there's no way I could get my wife to sign off on spending 5 grand (or more now) on a box with that much memory.
I own one, I don’t feel like the RAM is a huge issue (of course I want 192GB to run something like DS4 Flash). The lack of FP4 and slow memory bandwidth is rough. NVFP4 support is such a huge advantage that I would recommend others to buy a DGX spark over a strix halo if you’re using it purely for AI. Strix halo works better for general computing.
Heads up, you can absolutely run DS4 Flash on a 128gb machine - I have it running on my Strix Halo box right now.
https://github.com/antirez/ds4
How has it been and how’s the speed? I read online that the it’s ~200 TPS for PP and ~15 TPS for TG. Unfortunately for those speeds, it’s very very hard to use for agentic stuff.
That's pretty accurate to what I've seen, I'd definitely recommend a smaller model for active agentic use on that hardware.
It definitely seems to be the leader on 'general intelligence' on this hardware from my casual usage, but the newer Qwen or Gemma series models are much more usable speed-wise for agentic use, and often is as good or better than DS4 on that front.
There's a glimmer of hope with ROCmFP4 which seems to double the current throughput: https://github.com/charlie12345/rocmfp4-llama
I saw that but end of the day, the chips themselves don’t have hardware support for FP4. There’s smart ways around this limitation but it will never natively be close to true FP4 performance like MXFP4 and NVFP4 (happy to be proven wrong though).
The Strix Halo is a great dev machine and a mediocre AI machine. You can run Qwen 3.6 27B at a decent speed, or larger MoE models, and that's about it. For some that's more than enough though, myself included.
seems like an absolute amateur wrote this article
Until RAM prices drop and can economically get machines with 256GB, 512GB and higher bandwidth... I frankly think the local AI story is going to be still fairly muted for most people.
My Spark can do Qwen3.6 MoE A3B at 60 to 70-ish token/second and that's really good, but there's limits the usefulness of that model. It's not useful for coding, in any case.
Once people can run something like GLM 5.2 at lower quants (512GB could do a passable job), then I think the story changes.
Whether we ever see DRAM as cheap as it was ever again, I don't know.
Agree - the 128GB Strix Halo is capable if you use LLMs as assistants, but it's not so good if you use LLMs as agents (or worse, agent teams/swarms) since all of the models that can fit on it are pretty dumb compared to frontier or near-frontier models. You can at best hope for Sonnet-level capabilities.
That doesn't mean that local models are useless though! If Mythos/Sol is an ASI that threatens to take your job and turn you into paperclips, then Qwen/Gemma is an old-fashioned office secretary that loyally helps you with tasks but doesn't have a good grasp of details. Every white-collar worker 50 years ago would have killed to have a hard-working personal secretary.
Exactly this. I own the Framework desktop board. I knew all of its limitations before I bought it, and it's ok to play with on a hobby level, but it isn't much more than a Radio Shack toy.
That memory bandwidth is painful. It's like trying to fill an Olympic swimming pool with a thimble.
It is excellent as a regular PC, though.
Part of me wonders, would 3d xpoint (if still around) be a viable option?
Yeah it is slower than real RAM by a good amount for latency, but you can get similar bandwidth and the cost was history about half of the same size DDR.
Possibly - I've heard anecdotal reports that old Intel Optane chips are in hot demand right now. Intel/Micron probably would have made a killing if they had kept that product line alive for a few more years. Never miss an attempt to snatch defeat from the jaws of victory!
I was thought experimenting the other day... ~10 nVME drives striped and running parallel could approximate the memory bandwidth of DDR5 DRAM in a box like this. Like you say, latency wouldn't compete but on raw throughput would be comparable.
Not anymore cost effective, I guess, but gets you the ability to work over very large model sizes maybe. But the problem is that tensor matmul etc hardware wouldn't work effectively with it.
Useful for KVCache though.
I'll just leave this here: "Achieving 11M IOPS and 66 GB/S IO on a Single ThreadRipper Workstation" (2021), https://news.ycombinator.com/item?id=25956670 / https://tanelpoder.com/posts/11m-iops-with-10-ssds-on-amd-th...
We're maybe only 2 years away from really useful, relatively affordable local LLM usage. You can buy a 5090 PC for $5-6k but 32GB of VRAM really limits model sizes to ~31B. And that won't change (even with NVidia's next generation) because NVidia uses VARM as an aggressive market segmentation technique.
No, the hope really is these other platforms with a shared memory architecture. The DGX Spark won't be it because of the aforementioned market segmentation. So that leaves two players: AMD and Apple.
The AMD platform is still too low memory bandwidth, currently <300GB/s. For comparison a 5090 (or 6000 Pro) is 1.8TB/s and the M3 Ultra Mac Studio is ~900GB/s. Oh and B100/B200 uses HBM3e memory at ~3.2TB/s. The M5 Max in some Macbook Pros tops out at ~600GB/s. So you need access to better RAM and better CPU architecture for all this.
My great white hope is Apple. They have the market power to get memory and build silicon that coul dhave enough FLOPS to compete with NVidia's platform. They've started talking about it and I've seen rumors they're targeting the M7 generation (2028) for a huge leap. I'll believe it when I see it however.
But the point is, I think we'll be running 31B models at 100+tok/s on enthusiast hardware in 2 years and we'll likely be able to locally run 100-400B models, possibly larger.