Posted by low memory bandwidth
- Today at 10:03:36
While 128 GB RAM are nice, Panther Lake is a 128-bit APU and, as such, is approx. 2x slower than a Strix Halo system:
Strix Halo: 256 GB/s = 256-bit * 8000 MT/s / 1000 / 8.
MS-03 mini-PC: 115 GB/s = 128-bit * 7200 MT/s / 1000 / 8.
That being said, if you think that a 128 GB RAM Strix Halo is too expensive, know that it may not even be needed:
Qwen3.5-122B-A10B (MoE architecture), according to their own evaluations, is about as good as Qwen3.5-27B (dense architecture) (yes, Qwen3.6-27B been out a while ago, but no Qwen3.6-122B-A10B, I 'wonder' why): huggingface.co/Qwen/Qwen3.5-122B-A10B#benchmark-results. For the same amount of parameters, a dense model performs about 4 to 5 times better. MoE models are designed to run fast enough from RAM and dense models are designed for run from VRAM: 27B/A10B = 2.7 times speed difference, it's a speed vs size trade-off. If mini-PCs, like this, become so expensive, it starts to make sense to run a dense model again. All you need is a 24 GB VRAM, or more, GPU, to run a good enough quant (huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF UD-Q4_K_XL quant (17.9 GB)). Compare the price of such a GPU vs this mini-PC and decide.
128 GB RAM just don't compensate enough for a MoE vs dense model. What is really needed are at least 192 GB RAM, preferably 256 GB RAM, to run quants of models like DeepSeek V4 Flash (it's still just a preview in the following AA evaluation and a final version is supposed to arrive this month/Juli), MiMo-V2.5, MiniMax-M3. See evaluations:
artificialanalysis.ai/?models=qwen3-5-122b-a10b-non-reasoning%2Cqwen3-5-122b-a10b%2Cmimo-v2-5-0424%2Cqwen3-6-27b-non-reasoning%2Cminimax-m2-7%2Cqwen3-6-35b-a3b-non-reasoning%2Cstep-3-7-flash%2Cqwen3-6-35b-a3b%2Cdeepseek-v4-flash-high%2Cqwen3-6-27b%2Cdeepseek-v4-flash-non-reasoning%2Cgpt-oss-120b-low%2Cgemma-4-31b%2Cgemma-4-31b-non-reasoning%2Cdeepseek-v4-flash%2Cmistral-medium-3-5%2Cgpt-oss-120b%2Cminimax-m3&intelligence=artificial-analysis-intelligence-index#intelligence-tabs