TS

whatever1983

AI & ML interests

None yet

Recent Activity

commented on an article 12 days ago

Comparing sub 50GB Llama 4 Scout quants (KLD/Top P)

commented on an article 12 days ago

Comparing sub 50GB Llama 4 Scout quants (KLD/Top P)

new activity 4 months ago

bullerwins/DeepSeek-V3-GGUF:Please quantize base model too

View all activity

Organizations

None yet

whatever1983's activity

commented on Comparing sub 50GB Llama 4 Scout quants (KLD/Top P) 12 days ago

@bartowski :

mradermacher is having trouble with Maverick 128E IQ2_M. The embedding layer needs to be 3 bits. That's the whole difference between IQ2_M and IQ2_S. For coding, you need the embedding to be 3bits. Anything less messes up the coding(not that the llama4 models are good at coding, maybe an Athene or Dracarys finetune will bring coding up to par, or maybe Meta releases the 0236 version). Can you see if B5125 fixes the Maverick 128E IQ2_M?

commented on Comparing sub 50GB Llama 4 Scout quants (KLD/Top P) 12 days ago

skipped over IQ2M, Maverick 128E errored out on IQ2M. Don't know if B5125 fixed it. Anything under IQ2_M is basically garbage for coding regardless for PPL measurements.

New activity in bullerwins/DeepSeek-V3-GGUF 4 months ago

Please quantize base model too

#3 opened 4 months ago by

Delta36652

New activity in deepseek-ai/DeepSeek-V3 4 months ago

How can we thank you enough, whale bros?

#1 opened 4 months ago by

KrishnaKaasyap