On my SpacemiT K3 SBC with 32GB RAM (where models run on the eight A100 RISC-V cores with 1024 bit vectors) doing the same task I got 5, 5.8, 6.5 tok/s using gemma-4-26B-A4B-it-QAT-Q4_0.gguf, Qwen3.6-35B-A3B-Q4_K_M.gguf, Qwen3.5-35B-A3B-Q4_K_M.gguf. The corresponding dense models are more in the 2.5-3 tok/s range.
Kind of slow, but using only 14W of electricity so the Wh per task is twice as good as using my i9-13900 laptop with 4060 GPU.
I use it with a strix halo server. 35B runs stupidly fast. 27B is about 700 TPS prefill and 30 TPS token generation. Which interestedly is about what Kimi K3 gives me depending on provider.
While the laptop option is nice, for an inference server you're probably going to want the desktop form factor as it has significantly more thermal overhead and thus better performance. In the desktop models most of the internal volume is a gigantic heatsink
I have a framework desktop, but depending on your need, DGX spark might be better. The prefill and NVFP4 is a significant advantage. But framework desktop is a better general computer. I expect to be able to use it for years to come. Where as DGX Spark you’re at the mercy of NVIDIA BSP.
On an 8GB GPU and 32GB laptop: ~5 words/s while running in Qubes via ollama with completely default settings (I don't have an install at the moment that'll tell me tokens/s). Not exactly a highly tuned setup, but it's a ballpark at least :)
Tolerable and usable for some things, though thinking makes it take about a minute to reply in many cases. But getting this kind of thing to run on 8GB of VRAM is the main benefit of the mix-of-experts setup: it can do partial GPU loading and get a ton better throughput than a similarly-sized dense model (like 5-10x, sometimes more).
It's pretty fast, faster than I could type anyway, but not as fast as Claude of course. My oMLX dashboard says I get about 45 tokens per second from the Qwen model I'm running (I host it on my M1 Mac Studio, not on my laptop).
On my SpacemiT K3 SBC with 32GB RAM (where models run on the eight A100 RISC-V cores with 1024 bit vectors) doing the same task I got 5, 5.8, 6.5 tok/s using gemma-4-26B-A4B-it-QAT-Q4_0.gguf, Qwen3.6-35B-A3B-Q4_K_M.gguf, Qwen3.5-35B-A3B-Q4_K_M.gguf. The corresponding dense models are more in the 2.5-3 tok/s range.
Kind of slow, but using only 14W of electricity so the Wh per task is twice as good as using my i9-13900 laptop with 4060 GPU.
I use it with a strix halo server. 35B runs stupidly fast. 27B is about 700 TPS prefill and 30 TPS token generation. Which interestedly is about what Kimi K3 gives me depending on provider.
what hardware do you use or recommend for this? never heard of it until today.
Strix Halo is the unified memory platform from AMD. Similar to the DGX Spark from NVIDIA or the M series Macs.
I personally have the Framework Desktop, but there's also systems from other brands like Bosgame
You can also get it in a laptop form factor that feels like a MBP with a nicer keyboard if you get an HP Zbook G1A!
Huge fan of that thing, it's th e Linux MBP I've always wanted.
While the laptop option is nice, for an inference server you're probably going to want the desktop form factor as it has significantly more thermal overhead and thus better performance. In the desktop models most of the internal volume is a gigantic heatsink
I have a framework desktop, but depending on your need, DGX spark might be better. The prefill and NVFP4 is a significant advantage. But framework desktop is a better general computer. I expect to be able to use it for years to come. Where as DGX Spark you’re at the mercy of NVIDIA BSP.
RTX 4060 and above. Ideally RTX 50 Series, because you can run NVFP4-quantized GGUFs that give you better prefill AND better quality.
On an 8GB GPU and 32GB laptop: ~5 words/s while running in Qubes via ollama with completely default settings (I don't have an install at the moment that'll tell me tokens/s). Not exactly a highly tuned setup, but it's a ballpark at least :)
Tolerable and usable for some things, though thinking makes it take about a minute to reply in many cases. But getting this kind of thing to run on 8GB of VRAM is the main benefit of the mix-of-experts setup: it can do partial GPU loading and get a ton better throughput than a similarly-sized dense model (like 5-10x, sometimes more).
It's pretty fast, faster than I could type anyway, but not as fast as Claude of course. My oMLX dashboard says I get about 45 tokens per second from the Qwen model I'm running (I host it on my M1 Mac Studio, not on my laptop).
[dead]