toktapethe record, not a recording of it

midagedev

https://github.com/midagedev

20 runs, newest first

GLM-5.3-Flash graftB Q6_K · MTP · ik_llama.cpp 24.2 tok/s

GLM-5.3-Flash-graftB.gguf

ik_llama.cpp425a2c1d Q6_KNVIDIA RTX A6000 / NVIDIA GeForce RTX 3090linux 1 stream ! 2 caveats

2026-09-18

GLM-5.3-Flash EXL3 4.05 bpw · 2 streams · exllamav3 19.4 tok/s

GLM-5.3-Flash-exl3-4.05

exllamav31.5.0EXL3 4.05 bpw · head 6.0NVIDIA RTX A6000 / NVIDIA GeForce RTX 3090linux 2 streams ! 2 caveats

2026-09-18

GLM-5.3-Flash EXL3 4.05 bpw · 1 stream · exllamav3 22.0 tok/s

GLM-5.3-Flash-exl3-4.05

exllamav31.5.0EXL3 4.05 bpw · head 6.0NVIDIA RTX A6000 / NVIDIA GeForce RTX 3090linux 1 stream ! 3 caveats

2026-09-18

Qwen3-Coder-Next IQ4_XS · 1 stream · long answer 133 tok/s

Qwen3-Coder-Next-IQ4_XS.gguf

ik_llama.cppc10fbbcc IQ4_XSNVIDIA GeForce RTX 3090 / NVIDIA RTX A6000linux 1 stream ! 3 caveats

2026-09-18

Qwen3.8-Flash-Next UD-Q4_K_XL · 1 stream · mlock 51.0 tok/s

Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf

llama-serverb0UD-Q4_K_XLNVIDIA GeForce RTX 3090 / NVIDIA RTX A6000linux 1 stream

2026-09-18

DeepSeek-V4.1-Flash Q3_K_M · Korean prose, no thinking 20.3 tok/s

DeepSeek-V4.1-Flash-Q3_K_M-00001-of-00009.gguf

llama-serverb96 Q3_K_MNVIDIA RTX A6000 / NVIDIA GeForce RTX 3090linux 1 stream ! 2 caveats

2026-09-18

DeepSeek-V4.1-Flash Q3_K_M · 4 streams 27.4 tok/s

DeepSeek-V4.1-Flash-Q3_K_M-00001-of-00009.gguf

llama-serverb96 Q3_K_MNVIDIA RTX A6000 / NVIDIA GeForce RTX 3090linux 4 streams ! 1 caveat

2026-09-18

DeepSeek-V4.1-Flash Q3_K_M · 1 stream · 347 GB model on 72 GB of VRAM 28.6 tok/s

DeepSeek-V4.1-Flash-Q3_K_M-00001-of-00009.gguf

llama-serverb96 Q3_K_MNVIDIA RTX A6000 / NVIDIA GeForce RTX 3090linux 1 stream ! 1 caveat

2026-09-18

DeepSeek-V2-Lite-Chat Q3_K_M · 4 streams 211 tok/s

DeepSeek-V2-Lite-Chat.Q3_K_M.gguf

ik_llama.cppc10fbbcc Q3_K_M rtx-a6000 linux 4 streams ! 1 caveat

2026-09-18

DeepSeek-V2-Lite-Chat Q3_K_M · 1 stream 183 tok/s

DeepSeek-V2-Lite-Chat.Q3_K_M.gguf

ik_llama.cppc10fbbcc Q3_K_M rtx-a6000 linux 1 stream ! 1 caveat

2026-09-18

Qwen2.5-7B-Instruct Q3_K_M · 4 streams · dense control 196 tok/s

Qwen2.5-7B-Instruct-Q3_K_M.gguf

ik_llama.cppc10fbbcc Q3_K_M rtx-a6000 linux 4 streams ! 1 caveat

2026-09-18

Qwen2.5-7B-Instruct Q3_K_M · 1 stream · dense control 116 tok/s

Qwen2.5-7B-Instruct-Q3_K_M.gguf

ik_llama.cppc10fbbcc Q3_K_M rtx-a6000 linux 1 stream ! 2 caveats

2026-09-18

Qwen3.6-35B-A3B UD-Q4_K_XL · 1 stream 140 tok/s

Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf

ik_llama.cppc10fbbccUD-Q4_K_XLNVIDIA RTX A6000 / NVIDIA GeForce RTX 3090linux 1 stream ! 3 caveats

2026-09-18

Offload ladder · 10 layers on CPU 102 tok/s

Qwen3.6-35B-A3B-UD-Q6_K.gguf

ik_llama.cppc10fbbcc UD-Q6_K rtx-a6000 linux 1 stream ! 2 caveats

2026-09-18

Offload ladder · 0 layers on CPU 129 tok/s

Qwen3.6-35B-A3B-UD-Q6_K.gguf

ik_llama.cppc10fbbcc UD-Q6_K rtx-a6000 linux 1 stream ! 3 caveats

2026-09-18

The README hero: four code-review questions at once 144 tok/s

Qwen3.6-35B-A3B-UD-Q6_K.gguf

ik_llama.cppc10fbbcc UD-Q6_K rtx-a6000 linux 4 streams ! 1 caveat

2026-09-18

Qwen3.6-35B-A3B Q6_K · 8 streams on 4 slots · ik_llama.cpp 152 tok/s

Qwen3.6-35B-A3B-UD-Q6_K.gguf

ik_llama.cppc10fbbcc UD-Q6_K rtx-a6000 linux 8 streams ! 2 caveats

2026-09-18

Qwen3.6-35B-A3B Q6_K · 4 streams · ik_llama.cpp 146 tok/s

Qwen3.6-35B-A3B-UD-Q6_K.gguf

ik_llama.cppc10fbbcc UD-Q6_K rtx-a6000 linux 4 streams ! 2 caveats

2026-09-18

Qwen3.6-35B-A3B Q6_K · 2 streams · ik_llama.cpp 130 tok/s

Qwen3.6-35B-A3B-UD-Q6_K.gguf

ik_llama.cppc10fbbcc UD-Q6_K rtx-a6000 linux 2 streams ! 2 caveats

2026-09-18

Qwen3.6-35B-A3B Q6_K · 1 stream · ik_llama.cpp 129 tok/s

Qwen3.6-35B-A3B-UD-Q6_K.gguf

ik_llama.cppc10fbbcc UD-Q6_K rtx-a6000 linux 1 stream ! 3 caveats

2026-09-18