toktapethe record, not a recording of it

NVIDIA GeForce RTX 3090

79 tapesbloomery51 llama-server17 ik_llama.cpp7 openai3 exllamav31 qwen3.6-35b-a3b10 qwen2.5-7b-instruct1 qwen3-coder-next-512x2.5b1

Narrowed toos: linux ×Clear all

79 runs, newest first

llama.cpp · GLM-5.3-Flash UD-Q4_K_XL · A6000 alone · coding review 497 in / 2000 out · PR #27754 MTP draft, -ncmoe 38 20.8 tok/sprefill 104 tok/s

GLM-5.3-Flash-UD-Q4_K_XL-00001-of-00006.gguf

llama-serverb11036UD-Q4_K_XL321B · 23B activeP497 · G2000 · cache 0%ctx 4kfa offoffload partialrtx-3090rtx-a6000linux1 stream ! 1 caveat

midagedev

2026-10-02

llama.cpp · Qwen3.8 Flash Next UD-Q4_K_XL · RTX 3090 alone · coding review 538 in / 1500 out · -ncmoe 43, no draft 35.5 tok/sprefill 73.1 tok/s

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

llama-serverb11157UD-Q4_K_XL177B · 10B activeP538 · G1500 · cache 0%ctx 8kfa onoffload partialrtx-3090rtx-a6000linux1 stream ! 4 caveats

midagedev

2026-10-01

bloomery · Qwen3.8 Flash Next UD-Q4_K_XL · RTX 3090 alone · coding review 538 in / 1500 out · residency + MTP draft 71.1 tok/sprefill 389 tok/s

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

bloomery0.1.0 (b7e9f46b-dirty)q4_KP538 · G1500 · cache 0%ctx 8koffload partialrtx-3090rtx-a6000linux1 stream ! 1 caveat

midagedev

2026-10-01

bloomery · Qwen3.8 Flash Next UD-Q4_K_XL · A6000 alone · coding review 538 in / 1500 out · residency + MTP draft 86.4 tok/sprefill 574 tok/s

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

bloomery0.1.0 (b7e9f46b-dirty)q4_KP538 · G1500 · cache 0%ctx 8koffload partialrtx-3090rtx-a6000linux1 stream ! 1 caveat

midagedev

2026-10-01

llama.cpp · Qwen3.8 Flash Next UD-Q4_K_XL · A6000 + 3090 · coding review 538 in / 1500 out · -ncmoe 21, no draft 44.3 tok/sprefill 217 tok/s

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

llama-serverb11157UD-Q4_K_XL177B · 10B activeP538 · G1500 · cache 0%ctx 8kfa onoffload partialrtx-3090rtx-a6000linux1 stream ! 3 caveats

midagedev

2026-10-01

bloomery · V4.1 Flash Q3_K_M · 3090 + A6000 · coding review 507 in / 1500 out · residency + host streaming + DSpark 45.9 tok/sprefill 170 tok/s

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

bloomery0.1.0 (493ec88a)q3_KP507 · G1500 · cache 0%ctx 16koffload partialrtx-3090rtx-a6000linux1 stream

midagedev

2026-09-30

ik_llama.cpp · V4.1 Flash Q3_K_M · 3090 + A6000 · coding review 507 in / 1500 out · DSpark n_max=3 28.8 tok/sprefill 156 tok/s

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

ik_llama.cpp81af2772 Q3_K_M 748B · 16B activeP507 · G1500 · cache 0%ctx 16koffload partialrtx-3090rtx-a6000linux1 stream ! 2 caveats

midagedev

2026-09-30

bloomery · V4.1 Flash Q3_K_M · 3090 + A6000 · chat expert list + DSpark draft 46.4 tok/sprefill 124 tok/s

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

bloomery0.1.0 (a692575e)q3_KP29 · G400 · cache 0%ctx 16koffload partialrtx-3090rtx-a6000linux1 stream ! 1 caveat

midagedev

2026-09-29

bloomery · V4.1 Flash Q3_K_M · 3090 + A6000 · chat expert list + DSpark draft 43.8 tok/sprefill 90.6 tok/s

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

bloomery0.1.0 (a692575e)q3_KP26 · G400 · cache 0%ctx 16koffload partialrtx-3090rtx-a6000linux1 stream ! 2 caveats

midagedev

2026-09-29

MiMo-V2.6-Flash-RL-MXFP4-00001-of-00002.gguf 31.1 tok/sprefill 266 tok/s

llama-serverb0 MXFP4 309B · 20B activeP765 · G843 · cache 0%ctx 32kfa onoffload partialrtx-3090rtx-a6000linux2 streams ! 3 caveats

anonymous

2026-09-22

MiMo-V2.6-Flash-RL-MXFP4-00001-of-00002.gguf 24.0 tok/sprefill 293 tok/s

llama-serverb0 MXFP4 309B · 20B activeP6288 · G1646 · cache 0%ctx 32kfa onoffload partialrtx-3090rtx-a6000linux1 stream ! 3 caveats

anonymous

2026-09-22

GLM-5.3-Flash EXL3 4.05 bpw · 2 streams · exllamav3 18.4 tok/sprefill 308 tok/s

GLM-5.3-Flash-exl3-4.05

exllamav31.5.0EXL3 4.05 bpw · head 6.0313B · 21B activeP5296 · G548 · cache 0% · min 539ctx 16koffload partialrtx-3090rtx-a6000linux2 streams ! 5 caveats

midagedev

2026-09-20

Qwen3-Coder-Next IQ4_XS · 1 stream · long answer 129 tok/sprefill 1821 tok/s

qwen3-coder-next-512x2.5b

ik_llama.cpp7cd62a3e IQ4_XS 80B · 4.2B activeP6881 · G5061 · cache 0%ctx 32kfa onrtx-3090rtx-a6000linux1 stream ! 2 caveats

midagedev

2026-09-20

Qwen2.5-7B-Instruct Q3_K_M · 4 streams · dense control 134 tok/sprefill 4159 tok/s

qwen2.5-7b-instruct

ik_llama.cpp7cd62a3e Q3_K_M 7.6BP5665 · G934 · cache 0% · min 703ctx 8kfa onrtx-3090rtx-a6000linux4 streams ! 4 caveats

midagedev

2026-09-20

DeepSeek-V4.1-Flash Q3_K_M · 1 stream · 347 GB model on 72 GB of VRAM 23.8 tok/sprefill 52.3 tok/s

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

llama-serverb102 Q3_K_M 748B · 18B activeP6509 · G66 · cache 0%ctx 8koffload partialrtx-3090rtx-a6000linux1 stream ! 4 caveats

midagedev

2026-09-20

Offload ladder · 10 of 40 blocks' experts on CPU 97.9 tok/sprefill 1756 tok/s

qwen3.6-35b-a3b

ik_llama.cpp UD-Q6_K 35B · 3.5B activeP7246 · G2048 · cache 0%ctx 32kfa onoffload partialrtx-3090rtx-a6000linux1 stream ! 1 caveat

midagedev

2026-09-20

Qwen3.6-35B-A3B Q6_K · 8 streams on 8 slots · ik_llama.cpp 11.4 tok/sprefill 858 tok/s

qwen3.6-35b-a3b

ik_llama.cpp7cd62a3e UD-Q6_K 35B · 3.5B activeP5661 · G74 · cache 0% · min 64ctx 8kfa onrtx-3090rtx-a6000linux8 streams ! 5 caveats

midagedev

2026-09-20

Qwen3.6-35B-A3B Q6_K · 1 stream · ik_llama.cpp 125 tok/sprefill 2734 tok/s

qwen3.6-35b-a3b

ik_llama.cpp7cd62a3e UD-Q6_K 35B · 3.5B activeP7248 · G5271 · cache 0%ctx 32kfa onrtx-3090rtx-a6000linux1 stream ! 3 caveats

midagedev

2026-09-20

The README hero: four code-review questions at once 70.9 tok/sprefill 1097 tok/s

qwen3.6-35b-a3b

ik_llama.cpp7cd62a3e UD-Q6_K 35B · 3.5B activeP6065 · G603 · cache 0% · min 597ctx 8kfa onrtx-3090rtx-a6000linux4 streams ! 4 caveats

midagedev

2026-09-20

Published runs · every tape on this page was recorded on this card and carries its own caveats; the page does not rank them. toktape publish <run.toktape> puts one here.

Page views are counted with Cloudflare Web Analytics: no cookies, no fingerprinting. The toktape binary itself sends nothing anywhere.