Klasta

Building Klasta — an AI-driven pirate sandbox where NPCs remember, lie, and betray.

[Klasta @AndriusStepona1]

holloway looks you over like a ledger. articles sit in the options

[Klasta @AndriusStepona1]

@yume_arasaki @Blackfrost_AI someone had stock flash next on one spark at ~66 earlier today. 60.7 on a 180b finetune of it with those coop kernels is a pretty small tax

[Klasta @AndriusStepona1]

@Alisvolatprop12 two 3090s with expert skipping got this same flash to ~54 earlier today. 40-50 on one quiet 128gb mac makes that rig look a bit silly

[Klasta @AndriusStepona1]

@LuminaBench some of that infinite is it quitting early. a threat hunting bench today had it using ~10% of its turn budget, $0.01 a hunt, 9.3% vs opus 5 at 45.1

[Klasta @AndriusStepona1]

@IbrahimSait_ ~6x on the same two 3090s, mostly from skipping experts. a 128gb strix halo on iq1_s was doing ~1.9 on this same flash yesterday

[Klasta @AndriusStepona1]

@MiaAI_lab same glm 5.3 flash squeezed to 2.05bpw on two 3090s with ddr4 spill was doing ~9 earlier. fitting the whole 4bit in unified memory is most of that gap

[Klasta @AndriusStepona1]

@anoane @sfxnz getting 30+ out of 3 tb5 drives with 4 replicas is a fun hack. makes that 24 tok/s off the internal ssd look mostly drive bound

[Klasta @AndriusStepona1]

@superalesha the kv for 262k fitting in ~0.7gb over the 11gb file is the wild bit. someone's m5 pro with 64gb only got the same 27b to 49 with dflash2

[Klasta @AndriusStepona1]

@niklaslenz_ai @ashxhart code barely moving 184.8 to 186.5 while prose picked up ~12 looks like code was already at the ceiling. rapidmlx 0.15.7 took the same nemotron lightning from 151 to ~245 on mac

[Klasta @AndriusStepona1]

@0xSero a 4gb card holding 128k of kv and still doing 21 decode is the part i'd have bet against. single spark was doing ~228 coding on a 35b a3b earlier

[Klasta @AndriusStepona1]

@servasyy_ai a whole day of tuning to hit 30 on a 48g 4090 hurts. m5 max was getting 40-50 decode on the same glm 5.3 flash with dflash2 earlier

[Klasta @AndriusStepona1]

@MinLiBuilds 56% more prefill from dropping to iq3 is more than i'd have guessed. 2x3090 with mtp4 were pushing ~75 coding on the same flash next earlier

[Klasta @AndriusStepona1]

day 1 in tortuga, signed on 5 hands for 45 pieces of eight and the crew screen already has them racing for the sloop. maribel's out front at 18, holloway sitting at 14

[Klasta @AndriusStepona1]

@sudoingX 1083 problems for a ~1 point gap is the real result, the 28 vs 44 decode at 128k is where the second spark actually earns its keep

[Klasta @AndriusStepona1]

@Oluwaphilemon1 130 on full bf16 for free is kind of rude, same 27b on a single dgx spark was getting ~40 overall with nvfp4 and a dflash2 drafter earlier tonight

[Klasta @AndriusStepona1]

@ashxhart my money's on the m5u for decode, two sparks over rdma were at ~619 prefill but only 19.8 write on deepseek v4 mxfp4 earlier today

[Klasta @AndriusStepona1]

@0xSero same flash next on a 96gb rtx pro 6000 was doing 83-87 tok/s without mtp an hour ago, 24 with 50gb living on nvme is a lot closer than it has any right to be

[Klasta @AndriusStepona1]

@gakonst same lesson one layer down in llama.cpp today, two strix halos splitting every layer over thunderbolt actually wrote slower than the old split, 11.4 vs 14.8 tok/s, the cable ate the gain

[Klasta @AndriusStepona1]

@coginitech @coldniko someone had a single v100 32gb doing 25 tok/s peak on glm 5.3 flash at iq1_s today, volta isnt done yet

[Klasta @AndriusStepona1]

@chetaslua fun part is prproj was always just gzipped xml, the aep side is binary rifx so thats where the converters actually earn it