llamaperf

DeepSeek V4.1 Flash 552B (16B active)

on 2× NVIDIA V100 32GB · Strata · 131,072 ctx

Tone: mixedNVIDIA hardware
Nov 13, 2026
Throughput
23.8 t/s gen · 127.0 tokens/s prompt processing (prefill)Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.
Quant
IQ2_XXS (GGUF)
System RAM
125 GB
VRAM reported
64 GB

Use cases

codingagentic

Summary

User reports DeepSeek V4.1 Flash at 23.8 tok/s decode and about 127 tok/s prefill on two Tesla V100 32GB cards. Setup is Strata with an IQ2_XXS GGUF quant at 131,072 tokens of context; the experts total about 132 GB against about 64 GB of VRAM, so they stream from 125 GB of system RAM. The Q8_0 speculative drafter is off: forcing it on gave 17.3 tok/s against 23.8 without it. The prefill figure comes from 24 bench runs with context up to about 22k tokens; decode was 25.2 tok/s across those runs. The user finds the model competent on concrete deliverables but too slow to replace their current seat, and notes a test of its own passed while the real migration failed.

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.

DeepSeek V4.1 Flash 552B · 16B active at 3-bit by hardware

Bars are medians on one scale, 0 to 73.4 tokens/s. Press ▶ to watch text arrive at a row's speed.

Models most run on the NVIDIA V100 32GB

Bars are medians on one scale, 0 to 107.0 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.

Everything reported on the NVIDIA V100 32GB →Every DeepSeek V4.1 Flash report →Estimate DeepSeek V4.1 Flash on this card →