News:

Willkommen im Notebookcheck.com Forum! Hier können Sie über alle unsere Artikel und allgemein über notebookrelevante Dinge diskutieren. Viel Spass!

Main Menu

Framework launches world's first mini-ITX desktop PC with Ryzen AI Max+ Pro 495 and 192 GB RAM

Started by Redaktion, Yesterday at 22:16:22

Previous topic - Next topic

Redaktion

Framework has previewed the latest iteration of the Framework Desktop PC at AMD's Advancing AI event. The Framework Desktop is a small 4.1-litre mini-ITX PC that is powered by the Ryzen AI Max "Halo" APUs. The newest iteration brings the Ryzen AI Max+ Pro 495.

https://www.notebookcheck.net/Framework-launches-world-s-first-mini-ITX-desktop-PC-with-Ryzen-AI-Max-Pro-495-and-192-GB-RAM.1349336.0.html

A welcome change

Nice to see AMD strengthening their partner relationships besides the usual suspects like Asus.

Good news any why

Why this is good news and why 192, or better, 256 GB RAM is needed over 128 GB RAM:

For the same file-size and hence memory requirement, a dense AI LLM model outperforms a MoE AI LLM model by a lot, so a equally performing MoE model is multiple times larger, see 122B MoE vs 27B dense evaluation comparison: huggingface.co/Qwen/Qwen3.5-122B-A10B#benchmark-results, a difference of 122B/27B = 4.5 times in memory requirement. The advantage is that this 122B(-A10B) MoE model runs faster doe to fewer active parameters (27B vs 10B). And even larger ones, like 300B to maybe 450B, MoE models, run faster if they have fewer active parameters (e.g.: MiniMax-M3 (426B-23B)).

So, to significantly outperform a 27B dense in all metrics, a larger than 120B MoE model is needed.

Fortunately, there are many ~300 B MoE models to choose from. Here is a list of evaluated by AA (also see how good the much smaller 27B dense performs):
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%2Cnvidia-nemotron-3-super-120b-a12b%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%2Chy3%2Cgemma-4-31b%2Cgemma-4-31b-non-reasoning%2Cdeepseek-v4-flash%2Cmistral-medium-3-5%2Cgpt-oss-120b%2Cmotif-0714&intelligence=artificial-analysis-intelligence-index

Even a 120B model, at an 8-bit (quant), requires 120 GB of memory + context, so to fit these ~300B to ~400B models into 192 GB RAM, 2- to 5-bit quants are used.
Running ~400B MoE models at a 3-bit quant is also a possibility: E.g. MiniMax-M3 426B IQ3_XXS quant (159 GB).

Additionally, dense models are rare and the direction everyone seems to have chosen is MoE models.

Not evaluated yet by AA, because just released, but maybe later:
Solar-Open2 250B-A15B MoE
Laguna-S-2.1 - 120B-A8B MoE
More news about these models at reddit.com/r/LocalLLaMA.

AMD knows that more than 120B parameter models are needed and already exist and that their 2- to 5-bit quants perform well and this is why they announced Gorgon Halo (a Strix Halo rename/refresh, but with up to 192 GB RAM (160 GB allocatable on Windows, and maybe more on Linux, just like with Strix Halo)): notebookcheck.net/AMD-Ryzen-AI-Max-400-lineup-now-official-with-up-to-192-GB-RAM.1301777.0.html, which sees its (one of the) first implementation in this very news by FRAMEWORK.

A bit of recent history: For the same evaluation performance, MoE models' fewer active parameters were supposed to fix dense model's speed. RAM is typically much slower than VRAM, but also much larger than VRAM, so it was a perfect match. This is why MoE models came into being. But since recently, RAM prices have increased like 4 times and the advantage kinda went away (from this very news: ".. ends up costing over $4,000"). We need more dense models again, because memory prices are high.

PS: Most, if not all, desktop PC, mid-range B850 motherboards, support up to 256 GB RAM, but their memory bandwidth is less than half of this Halo product. Still, you can and should then also add 1 or even more (high VRAM) GPUs. This will compensate for the slower RAM by adding much faster VRAM and you end up having more total memory, too.

DS4 backing in llama.cpp

Meanwhile DS4 support is advancing in llama.cpp:
git log --pretty=format:"%h %ad %s" --date=short -P --grep "deepseek(?:.*?[^[:alnum:]])?v?4\b|ds[ -]?v?4(\.[0-9]+)*" -i
e8e6c7af2 2026-07-22 minor: fix reasoning preserve var for DS4 [no ci] (#25999)
f534da26e 2026-07-22 Fix DeepSeek4 crafted template (#25414)
67b9b0e7f 2026-07-22 llama-arch: fix DeepSeek4 APE tensor op (#25945)
846e991ec 2026-07-22 cuda: add sqrt_softplus in topk-moe for dsv4 (#25896)
91d2fc387 2026-07-20 llama_dsv4: write only used rows in state (#25325)
4937ca83f 2026-07-18 llama-quant : exclude i32 ffn_gate_tid2eid routing table from quantization (#25787)
5d5306bf3 2026-07-17 tests : initialize all tensors in test_dsv4_hc to avoid NaNs in sentinel tensors (#25822)
0dc74e332 2026-07-17 DeepseekV4: Add fused hyper-connection ops (#25585)
33a75f41c 2026-07-15 DeepseekV4: reduce graph splits (#25702)
7f575c39d 2026-07-14 DeepseekV4: fix seq_rm (#25588)
13f2b28b0 2026-07-11 DeepseekV4: clear cache only for seq rather than full (#25521)
00f5442cc 2026-07-11 ggml : add GGML_OP_LIGHTNING_INDEXER that implements DeepSeek V3.2/V4 lightning indexer (#24231)
2ed3c1abb 2026-07-10 llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k repeats in DeepSeek V4 (#25370)
024c46ae4 2026-07-07 llama: fix quantized kv-cache for dsv4 (#25202)
8c146a836 2026-06-29 DeepSeek V4  (#24162)
DeepSeek V4 Flash (huggingface.co/unsloth/DeepSeek-V4-Flash-GGUF) will fit nicely in this's 192 GB RAM and on Linux you should be able to use the full weights / 8-bit (162 GB) file.

DS4 backing in llama.cpp

To run AI models locally and privately see:
notebookcheck.net/Your-own-ChatGPT-offline-AI-without-the-cloud-on-your-laptop.1341872.0.html

Quick Reply

Name:
Email:
Verification:
Please leave this box empty:
Shortcuts: ALT+S post or ALT+P preview