Briskspirit

[Briskspirit @igorbiletski]

@___Harald___ And then Nvidia will make scary black assault GPU and it will get banned! And only DoD will be able to use it as it wasn’t initially listed in amendment 😅

[Briskspirit @igorbiletski]

@blind_via .357 Mag?

[Briskspirit @igorbiletski]

@___Harald___ Try MTP if you haven’t yet!) especially with 27B dense model

[Briskspirit @igorbiletski]

@aalleexx @___Harald___ @DanielW_Kiwi @__tinygrad__ That will be fixed :)

[Briskspirit @igorbiletski]

@blind_via Just a keyboard matrix, nothing fancy. But pcb looks great.

[Briskspirit @igorbiletski]

@blind_via @CapacitAnas I just ask ChatGPT to find it and it finds it))

[Briskspirit @igorbiletski]

@blind_via Not much cheaper.. but here you go:
https://t.co/9tJBJNoxW3
https://t.co/ruTUO3NCub
Smaller 1.27 pitch:
https://t.co/DdTo6CxD1c

[Briskspirit @igorbiletski]

@Vicharak_In Have you tried to check your links? From google and GitHub. You just lost me as an impulse buyer unfortunately.. https://t.co/kJLkfuL75m

[Briskspirit @igorbiletski]

@blind_via Don’t…

[Briskspirit @igorbiletski]

@ivanfioravanti Noticed that MLX for Qwen3.6 is “smarter” than unsloth gguf . How is it in your experience? Idk why but in my local benchmark 3.6 performs on par with 3.5 (both unsloth)..

[Briskspirit @igorbiletski]

@svgdevop @mudler_it @UnslothAI I am also not very happy… 3.6 performs worse than 3.5 on my mac with llama.cpp and MLX beats it (though most of the time gguf is “smarter”)

[Briskspirit @igorbiletski]

@jun_song It’s M5 Max 128Gb, tried both of your models - gguf with llama.cpp and mlx with mlx-vlm, also tried previous non multi modal with mlx-lm.

[Briskspirit @igorbiletski]

Getting more and more annoyed by people who scream “SOTA this, SOTA that”… Guys, have you even tried to test those models? For example getting my feed flooded with supergemma 4 and how it is cool and performs better than original model in every aspect. Really?? Nothing personal.

[Briskspirit @igorbiletski]

@Yampeleg Have you actually tested his model? My benchmark showed pretty bad results..

[Briskspirit @igorbiletski]

@jun_song So GGUF model is a lot worse than similar 4 bit original (don’t have it in Q4). And MLX model looks better but tool calling is totally broken!

[Briskspirit @igorbiletski]

@jun_song Any plans to support vision? Downloading, will test with my benchmark

[Briskspirit @igorbiletski]

@___Harald___ Skills, RAGs, tools… that’s a huge chunk of what makes Claude Code and Codex so nice. And temperature, penalties etc. Try to ask the same question to ChatGPT and Codex with the same model or Claude Code and Claude Chat.