bestmodel.run / planejar — agent view
rig: 4× B300 288GB · context 212000 tokens · basis: modeled (NOT measured)

model           quant         fits  need/usable GB   tok/s mid (low-high)   +MTP mid (low-high)   TTFT s
GLM-5.3 (full)  BF16          no    1516.1/1094.4    -                      -                     -
GLM-5.3 (full)  FP8 (official yes   768.7/1094.4     112 (84–157)           220 (133–371)         3.2
GLM-5.3 (full)  NVFP4         yes   433.6/1094.4     122 (90–177)           240 (142–420)         3.2
GLM-5.3 (full)  EXL3 3.5 bpw  yes   339.5/1094.4     125 (92–184)           247 (145–436)         3.2
GLM-5.3 (full)  GGUF Q4_K_M   yes   466.6/1094.4     88 (66–104)            174 (105–247)         3.2
GLM-5.3-Flash   BF16          yes   641.6/1094.4     175 (136–237)          343 (213–557)         1.3
GLM-5.3-Flash   FP8 (official yes   324/1094.4       202 (151–288)          395 (236–677)         1.3
GLM-5.3-Flash   NVFP4         yes   181.6/1094.4     216 (159–319)          423 (249–749)         1.3
GLM-5.3-Flash   EXL3 3.5 bpw  yes   141.6/1094.4     221 (161–329)          432 (253–772)         1.3
GLM-5.3-Flash   GGUF Q4_K_M   yes   195.6/1094.4     166 (122–199)          325 (191–467)         1.3

Method: t_step = active weight bytes / (BW x 0.75 x GPUs) + KV read / BW + layers x per-layer cost.
Per-layer cost band fitted on pool cells; see /blog/glm-5-3-vs-flash-4x-b300.