GLM 5.2 on 4x Sparks reasonable?

<!-- SC_OFF --><div class="md"><p>Every output token the model generates requires reading the entire KV cache from High Bandwidth Memory.</p> <p>Here's what that actually looks

<!-- SC_OFF --><div class="md"><p><a

<!-- SC_OFF --><div class="md"><p>Running Llama-3-70B with long contexts and hitting HBM limits constantly. 128K context burns approx 40GB just for KV cache — half the H100 gone

<!-- SC_OFF --><div class="md"><p>Got Qwen 3.5 27B to 1,103,941 tok/s on 12 nodes with 96 B200 GPUs.</p> <p>At that rate you process 50,000 insurance policy documents in hours

<!-- SC_OFF --><div class="md"><p>Currently using glm 4.7 flash, it’s very meh </p> <p>Heard omnicoder or Crow 9b are good, are they any better?</p> <p>Or Qwen3.5 27b?</p>

<!-- SC_OFF --><div class="md"><p>I could fit the Q3 model not sure if it's worth over 27B ? </p> </div><!-- SC_ON --> &#32; submitted by &#32; <a

<!-- SC_OFF --><div class="md"><p>I'm getting ~13 tps on Q8_0, with a context window of 128000, K Q8_0, V Q8_0</p> <p>this is on 3x GPUS (1x2060super 8gb, 2x5060ti 16gb), via

<!-- SC_OFF --><div class="md"><p>So i recently updated LMstudio after a long pause and updated my llama.cpp runtimes too.. i was shocked.. i thought maybe something like