GLM-5.3 320B (18B active) Flash
on 2× NVIDIA RTX 4060 Ti 16GB · Strata · 131,072 ctx
Summary
User reports GLM-5.3-Flash at a mean 13.0 tokens/s decode on 2x RTX 4060 Ti 16GB, with prefill of 165-166 tokens/s at 2048 and 385-386 tokens/s at 8192.
Setup is Strata v1.1.42 built from source with the unsloth UD-IQ3_XXS GGUF (4 shards) at 131072 context, expert cache auto and KV streaming on, model on a SATA SSD.
With STRATA_GLM_DEFER_CPU=1 decode was 13.4 and 13.7 tokens/s; with =2 (a separate PR) it was 14.0 tokens/s on both starts.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
GLM-5.3 320B · 18B active at 4-bit by hardware
- 37.4 t/s · one run
Bars are medians on one scale, 0 to 37.4 tokens/s. Press ▶ to watch text arrive at a row's speed.
Models most run on the NVIDIA RTX 4060 Ti 16GB
- Qwen3.8 27B3-bit20.0 t/s · one run
Bars are medians on one scale, 0 to 20.0 tokens/s. Press ▶ to watch text arrive at a row's speed.
Everything reported on the NVIDIA RTX 4060 Ti 16GB →Every GLM-5.3 report →Estimate GLM-5.3 on this card →