GGUF Q8 Performance

July 25, 2026 ยท View on GitHub

This report compares tested 16-bit GGUF packages against Q8_0 GGUF packages on CUDA. In the measured routes, Q8_0 improves wall time by up to 1.53x and lowers peak VRAM by up to about 37% compared with the matching 16-bit GGUF package.

Notes:

  • Offline session rows exclude the warmup request.
  • "Speed vs real time" means generated or processed audio duration divided by wall time.
  • "Q8 vs 16-bit" means 16-bit wall time divided by Q8 wall time; values above 1.0x mean Q8 was faster.
  • qwen3_tts uses the regenerated q8_v2 GGUF, which keeps speaker-sensitive tensors in 16-bit storage.
  • higgs_audio_tts 16-bit long-form uses text_chunk_size=256.

Summary

Q8 gives the clearest end-to-end wins on the larger AR-style models:

  • higgs_audio_tts: Q8 is 1.38x-1.53x faster on warmed requests and lowers peak VRAM from 9.1 GiB to 5.9 GiB.
  • fish_audio: Q8 is 1.26x-1.34x faster on warmed requests and lowers peak VRAM from 11.8 GiB to 7.5 GiB.
  • voxtral_realtime: Q8 is 1.31x-1.38x faster in offline ASR and lowers peak VRAM from 10.7 GiB to 7.6 GiB.

Smaller or already memory-light models still load and run with Q8, but the speed gain can be modest. Treat Q8 as a measured route choice, not a guaranteed win for every model.

Voxtral Q4_K Quick Check

Voxtral also has a q4_k GGUF package. In this quick CUDA path-test check, Q4_K was smaller and faster than Q8_0 while producing effectively the same transcripts. Four requests matched exactly; one differed only by capitalization (Mother Nature vs mother nature).

RouteQ8 RTFQ4_K RTFQ4_K vs Q8
Offline short0.08620.06291.37x
Offline medium0.06430.04761.35x
Offline longer0.05760.04391.31x
Offline sampled0.06300.05001.26x
Streaming path0.10360.09041.15x

TTS Offline Long-Lived Session

Model16-bit speed vs real timeQ8 speed vs real timeQ8 vs 16-bit16-bit peak VRAMQ8 peak VRAM
pocket_tts68.5x-99.0x78.7x-102.0x1.01x-1.21x2060 MiB1932 MiB
chatterbox6.2x-8.2x6.2x-8.7x1.01x-1.11x4333 MiB4309 MiB
omnivoice7.2x-24.4x8.2x-29.0x1.08x-1.22x3250 MiB3157 MiB
qwen3_tts5.3x-7.0x5.8x-7.9x1.01x-1.12x7873 MiB6397 MiB
fish_audio2.5x-2.6x3.1x-3.4x1.26x-1.34x12093 MiB7669 MiB
higgs_audio_tts6.1x-6.7x8.8x-10.1x1.38x-1.53x9326 MiB6024 MiB

TTS Offline Long-Form

Model16-bit speed vs real timeQ8 speed vs real timeQ8 vs 16-bit16-bit peak VRAMQ8 peak VRAM
pocket_tts82.6x85.5x1.04x2098 MiB2298 MiB
chatterbox7.9x8.5x1.06x5033 MiB4554 MiB
omnivoice38.6x43.7x1.13x3175 MiB3069 MiB
qwen3_tts5.9x6.6x1.12x9412 MiB8138 MiB
fish_audio2.5x3.3x1.27x12261 MiB9228 MiB
higgs_audio_tts6.1x8.5x1.41x11878 MiB9129 MiB

ASR Offline Long-Lived Session

Model16-bit speed vs real timeQ8 speed vs real timeQ8 vs 16-bit16-bit peak VRAMQ8 peak VRAM
voxtral_realtime11.1x-12.5x14.7x-16.7x1.31x-1.38x10909 MiB7754 MiB
nemotron_asr277.8x-384.6x285.7x-400.0x1.03x-1.12x5125 MiB4028 MiB

ASR Streaming Long Audio

Model16-bit server TTFTQ8 server TTFT16-bit client TTFTQ8 client TTFT16-bit speed vs real timeQ8 speed vs real time16-bit peak VRAMQ8 peak VRAM
voxtral_realtime207.308 ms179.896 ms550.526 ms530.558 ms4.7x5.4x12616 MiB8972 MiB
nemotron_asr205.007 ms214.822 ms488.629 ms499.453 ms31.6x33.4x2816 MiB2497 MiB