The following text is about running AI locally (and privately) on these devices.Quote24 GB of memory [..] costs $2,999
24 GB of unified memory / RAM / total memory (24 GB RAM + 0 GB VRAM / no dedicated GPU) is not enough memory for general purpose SOTA AI models, like Qwen3.8-27B dense, Qwen3.6-35B MoE or any other models in a similar size and even 32 GB is very limited: 32 GB unified memory / RAM of still SOTA (in its class size) (or any other models in similar 35B MoE class size) (and there's still no smaller MoE model that outperforms it on average) AI LLM model[1] huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF at a minimum good Q4_K_XL quant, with proper context, fit about this much of context tokens:
Quote from: reddit.com/r/LocalLLaMA/comments/1sq94qx/is_anyone_getting_real_coding_work_done_with.. I've come to the conclusion that (1) 32768 is the biggest context I can get away with in an adequately smart model, and (2) it just ain't enough.
It's also possible to calculate the available context[2]: 16,384 [KV cache context tokens per GB]*(32 [GB total memory] - 6 to 8 [GB for the OS] - [GB quant filesize]):
~34,400 context tokens = 16,384*(32-7[3]-22.9) (confirms the quote)
~18,600 context tokens = 16,384*(32-8-22.9) (1 GB more for the OS, this is how much every GB matters)
Recalculating with an additional
8 GB VRAM GPU of a normal gaming laptop:
165,000 context tokens = 16,384*(32-7-22.9+
8)
So, in this memory sense, any gaming laptop that has a 6-8 GB VRAM GPU is superior and maybe even cheaper (700-800 bucks for 16 GB RAM config, then just upgrade to 32 GB RAM) and it's also going to have upgradable RAM. The additional 8 GB of (fast) memory/VRAM of the GPU make all the difference in being able to fit and run these general purpose SOTA AI models at a useful context size.
Qwen4-27B has been announced[4] (and is rumored to be released by the end of october) and hopefully it's MoE arch as well, not only dense arch. because then it will run fast on 32 GB unified memory devices, and this will revive them:Based on from what we know from Qwen3.8-Flash-Next (Qwen4 preview), such a model could be roughly 20-30% smaller and still perform better. Smaller means the 30% would be streamed from the SSD with only a ~5% performance penalty vs if it was streamed from the RAM[6].
Recalculating context from existing 27B Q4_K_XL (17.6 GB) quants:
121,000 = 16,384*(32-7-17.6).
Including a possible reduction in KV cache context tokens memory requirements, like in [5], and even doubling the context tokens per GB, would double the context from 121,000 to 242,000 tokens!
If it doesn't have to be a notebook: You get much more AI performance (=memory capacity + GPU compute/performance) per buck in a classical desktop PC.
[1] artificialanalysis.ai/models/open-source/small?models=qwen3-8-27b-medium%2Cqwen3-8-27b-low%2Ck2-horizon-mova-36b-a4b%2Cg9v3-39a5b%2Ck2-horizon-7b%2Cqwen3-8-27b-non-reasoning%2Cqwen3-6-35b-a3b%2Cmuse-glimmer%2Cgemma-4-26b-a4b%2Cqwen3-6-35b-a3b-non-reasoning%2Cgemma-4-31b%2Cqwen3-8-27b#artificial-analysis-intelligence-index (K2 Horizon has expensive KV cache and may reason above average[6])
[2] reddit.com/r/Qwen_AI/comments/1vo8pjz/qwen3827b_kv_cache_works_out_to_64_kibtoken_so/
[3] techpowerup.com/353395/windows-11-26h2-update-lowers-ram-consumption[/quote]
[4] reddit.com/r/LocalLLaMA/comments/1wmxfjs/qwen_4_announced_at_apsara_conference/
[5] huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash (see "Global KV Cache Per Token (Bytes)" image)
[6] reddit.com/r/LocalLLaMA/comments/1wg0vqz/comment/p9qj8c8/
Surface Ultra - beautiful hardware, cursed by the software.. And based on previous experience with Surface, there will be no easy way (if at all) to install the proper operating system there.