JP

Eclaireur

[JP @jppol]

@analogalok And I didn't want to Claude to access my files/stuff directly (even if it could still via mcp)

[JP @jppol]

@analogalok and when the students are taught well, you have an autonomous local team for free, and the frontier model as a supervisor with few token spent.
I struggled too much with the teaching part of local models, so I came up with that.

[JP @jppol]

@analogalok Hey, thanks for the test results! Playing on my spare time with Hermes Agent and Qwen 3.5 9B, most reliable so far for me but still struggling with params. Need to try again Gemma4. I've built a bridge to have Claude talk to my local model and teach him skills. Next step is RAG.

[JP @jppol]

@analogalok Are you using Tomtom fork for llama.cpp turboquant? Could not see any difference so far.
28 t/s on 8 VRAM.
gemma-4-26B-A4B-
it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp --spec-draft-model mtp-gemma-
4-26B-A4B-it.gguf --spec-draft-n-max 6 --spec-draft-p-min 0.7 -c 80000 -v

[JP @jppol]

@WhaleEverything @NoLimitGains @grok GOLD

[JP @jppol]

@theUMreal @grok @NoLimitGains 👀

[JP @jppol]

@chydorina @remissionbiome @threadreaderapp unroll

[JP @jppol]

@silvano_trotta Combien ça coûte d'avoir une pièce d'identité aux États-Unis ? Le coût ne serait pas un frein également s'il est trop élevé ?

[Eclaireur @devenezleheros]

Ne gaspillez pas votre précieuse expérience pour montrer à quelqu'un qu'il a tort.

[JP @jppol]

@SecretsOfCrypto @AlphrFinance Good luck to all!