Nikolaos Perrakis

wannabe computer wizard

[Nikolaos Perrakis @Lalroch]

@jun_song Why jump from ~30b to ~400b? 120b like qwen3.5-122b are a good tier on their own.

[Nikolaos Perrakis @Lalroch]

@bcherny Fedora please. Coding is solved 😜

[Nikolaos Perrakis @Lalroch]

@LinuxHandbook Fedora 🐧

[Nikolaos Perrakis @Lalroch]

@BarathAnandan7 I m getting ~30t/s on Framework desktop at llama w/Vulkan. YMMV.

[Nikolaos Perrakis @Lalroch]

@ornith_ @atomic_chat_hq @ollama @UnslothAI @huggingface Awesome release, will test it locally! Congratulations to the team!
P.S. In the future also consider 122b-a10b architecture. With mtp of runs well on strix halo and dgx spark. It's a good step between 35/31B and 397B.

[Nikolaos Perrakis @Lalroch]

@UnslothAI Qwen-3.5-122b-a10b@UDQ4_K_XL MTP GGUF on @FrameworkPuter desktop
Also thanks, for making the mtp quants :)

[Nikolaos Perrakis @Lalroch]

@udiWertheimer Qwen-3.5-122b-a10b@UD-Q4_K_XL runs at around 30t/s with MTP at strix halo 128gb. Just saying...

[Nikolaos Perrakis @Lalroch]

@Shinnzo_xd And he can't even use Fable :p

[Nikolaos Perrakis @Lalroch]

@jun_song I m all for open source AI but let's be more careful. Mythos private release was April 7. Let's see how fast we get an open source model with the same capabilities...

[Nikolaos Perrakis @Lalroch]

@andrewqu Opus 4.8 treats anything not forbidden as allowed which can often be bad while gpt mostly treats anything not allowed as forbidden. Fable while constructing a plan for a change would say, oh by the way I found a bug there. Other models rarely if ever did that. OpenCode harness.

[Nikolaos Perrakis @Lalroch]

@jun_song For 128 GB Framework desktop what better model can you suggest compared to qwen3.6-35b-a3b@Q8 or qwen3.5-122b-a10b@Q4 ??

Using llama.cpp with vulkan for inference.

[Nikolaos Perrakis @Lalroch]

@thdxr It's too complex to remember. I ask Mythos when I need it.

[Nikolaos Perrakis @Lalroch]

@jun_song While I 'm all for that as a consumer the money are in enterprise inference. That's how Anthropic got ahead of OpenAI.

[Nikolaos Perrakis @Lalroch]

@thdxr (OpenCode Desktop) Even on a big screen if the model asks a question with many or elaborate possible answers there is usually not enough vertical space to comfortably read the text before the question. A minimize/restore of the question would help.

[Nikolaos Perrakis @Lalroch]

@topjohnwu Social media disproportionately amplifies negative voices. There are many of us who appreciate what you 've done. Many of us also lived enough to know we are not perfect at everything and need to experiment to learn new things. Ignore negativity and enjoy your family 🏡

[Son Luong @sluongng]

Codex just found a “workaround” of not having sudo on my pc…

[Nikolaos Perrakis @Lalroch]

@julien_c What's the % of difference in genes between humans and chimpanzees ?
;)

[Nikolaos Perrakis @Lalroch]

@MeKulukulu @ChristosPorios Το εχω ήδη κάνει.

[Nikolaos Perrakis @Lalroch]

@MeKulukulu @ChristosPorios Μπορείς να βάλεις το ίδιο το chargpt να κάνει review και κάτι τέτοιο θα το έβρισκε εύκολα.