mannix/ ornith-1.5-27b-a3b-coder:IQ4_NL

317 1 month ago

Code-targeted 184-expert prune of Ornith-1.5-35B-A3B (256e→184e, ~27B, A3B active, top-8). Beats the 256-expert teacher on LiveCodeBench v6 (+6.5pp) and HumanEval+ (+2.4pp) at ~28% fewer experts. Native MTP head + vision tower

vision tools thinking
... /
params
d75e3524b9de · 132B
{
"draft_num_predict": 3,
"min_p": 0,
"num_ctx": 32768,
"presence_penalty": 1.5,
"repeat_penalty": 1,
"temperature": 1,
"top_k": 20,
"top_p": 0.95
}