#1 · Cross-device average · llama.cpp · MTP
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
82.0Tater score
90.8% accuracy91.00 tok/s0.59s TTFT10.8 GiB peak RSS2 devices4 runs
Chat100.00%Routing89.88%Spudex90.00%Synthesis90.00%Tool Accuracy90.64%
See 4 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:04:13.758873+00:00 | 75.7 | 91.8% | 108.61 | 0.52s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:43:16.924204+00:00 | 88.3 | 89.9% | 73.35 | 0.67s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:17:07.547513+00:00 | 88.3 | 89.9% | 73.27 | 0.67s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:23:20.647032+00:00 | 75.7 | 91.8% | 108.76 | 0.52s |
#2 · Cross-device average · llama.cpp · MTP
TaterTotterson/Qwen3.6-35B-A3B-MTP-GGUF-Tater-NoThink
Qwen3.6-35B-A3B-UD-Q4_K_M.gguf
80.2Tater score
87.9% accuracy88.80 tok/s0.70s TTFT12.4 GiB peak RSS2 devices4 runs
Chat100.00%Routing84.88%Spudex84.16%Synthesis100.00%Tool Accuracy87.43%
See 4 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:07:29.363767+00:00 | 71.1 | 87.0% | 96.80 | 0.55s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:48:25.786706+00:00 | 89.0 | 88.8% | 78.42 | 0.85s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:22:15.434255+00:00 | 89.1 | 88.8% | 78.90 | 0.85s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:26:32.994760+00:00 | 71.5 | 87.0% | 101.07 | 0.55s |
#3 · Cross-device average · llama.cpp · DFLASH
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
74.2Tater score
90.8% accuracy39.78 tok/s0.62s TTFT12.9 GiB peak RSS2 devices4 runs
Chat100.00%Routing89.88%Spudex90.00%Synthesis90.00%Tool Accuracy90.64%
See 4 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:04:56.554299+00:00 | 68.9 | 91.8% | 36.20 | 0.52s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:44:15.074009+00:00 | 79.4 | 89.9% | 43.13 | 0.72s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:18:05.387711+00:00 | 79.5 | 89.9% | 43.37 | 0.72s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:24:03.026657+00:00 | 68.9 | 91.8% | 36.41 | 0.53s |
#4 · Cross-device average · mlx · BASELINE
TaterTotterson/Gemma-4-26B-A4B-IT-UD-Q4_K_XL-mlx-Tater-NoThink
674e8a504329b470c761c2e1feeb31d93d6ceeb7
73.6Tater score
95.3% accuracy49.77 tok/s0.26s TTFT18.5 GiB peak RSS1 device2 runs
Chat90.00%Routing91.88%Spudex96.67%Synthesis90.00%Tool Accuracy100.00%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:14:59.452371+00:00 | 73.6 | 95.3% | 49.54 | 0.26s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:33:45.435853+00:00 | 73.6 | 95.3% | 50.01 | 0.27s |
#5 · Cross-device average · llama.cpp · BASELINE
TaterTotterson/Qwen3.6-35B-A3B-MTP-GGUF-Tater-NoThink
Qwen3.6-35B-A3B-UD-Q4_K_M.gguf
67.7Tater score
75.8% accuracy65.65 tok/s0.67s TTFT11.2 GiB peak RSS2 devices4 runs
Chat100.00%Routing73.21%Spudex45.00%Synthesis67.50%Tool Accuracy87.43%
See 4 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:06:44.120075+00:00 | 69.4 | 87.0% | 78.22 | 0.53s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:47:02.097192+00:00 | 65.9 | 64.7% | 52.47 | 0.82s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:20:51.859924+00:00 | 65.9 | 64.7% | 52.58 | 0.82s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:25:48.331279+00:00 | 69.5 | 87.0% | 79.31 | 0.53s |
#6 · Cross-device average · llama.cpp · MTP
TaterTotterson/Qwen3.8-27B-GGUF-Tater-NoThink
Qwen3.8-27B-Tater-NoThink-Q4_K_M.gguf
67.6Tater score
88.8% accuracy22.82 tok/s2.58s TTFT14.0 GiB peak RSS2 devices4 runs
Chat100.00%Routing85.29%Spudex90.00%Synthesis100.00%Tool Accuracy87.43%
See 4 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:14:11.796656+00:00 | 64.9 | 88.8% | 24.75 | 2.82s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:55:28.279842+00:00 | 70.3 | 88.8% | 21.20 | 2.35s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:29:18.938662+00:00 | 70.2 | 88.8% | 21.17 | 2.34s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:33:10.300966+00:00 | 64.9 | 88.8% | 24.18 | 2.79s |
#7 · Cross-device average · llama.cpp · BASELINE
unsloth/Qwen3.5-0.8B-GGUF
Qwen3.5-0.8B-Q4_K_M.gguf
67.2Tater score
53.1% accuracy214.21 tok/s0.10s TTFT1.2 GiB peak RSS1 device1 run
Chat100.00%Routing35.12%Spudex36.67%Synthesis90.83%Tool Accuracy60.05%
See 1 individual run
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:14:21.901105+00:00 | 67.2 | 53.1% | 214.21 | 0.10s |
#8 · Cross-device average · llama.cpp · BASELINE
TaterTotterson/Qwen3.8-27B-GGUF-Tater-NoThink
Qwen3.8-27B-Tater-NoThink-Q4_K_M.gguf
61.4Tater score
81.0% accuracy18.91 tok/s2.54s TTFT10.7 GiB peak RSS2 devices4 runs
Chat100.00%Routing83.78%Spudex45.00%Synthesis95.00%Tool Accuracy86.00%
See 4 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:10:50.330144+00:00 | 65.2 | 88.8% | 26.61 | 2.78s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:52:11.229193+00:00 | 57.5 | 73.1% | 10.96 | 2.32s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:26:01.254512+00:00 | 57.6 | 73.3% | 10.97 | 2.32s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:29:50.869875+00:00 | 65.2 | 88.8% | 27.08 | 2.73s |
#9 · Cross-device average · llama.cpp · BASELINE
TaterTotterson/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF-Tater-NoThink
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Tater-NoThink-Q4_0.gguf
60.1Tater score
61.3% accuracy80.31 tok/s0.72s TTFT9.2 GiB peak RSS2 devices4 runs
Chat86.25%Routing40.77%Spudex39.16%Synthesis90.00%Tool Accuracy78.15%
See 4 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:05:48.227024+00:00 | 70.2 | 85.1% | 100.08 | 0.57s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:45:37.242320+00:00 | 49.8 | 37.4% | 59.27 | 0.87s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:19:27.390900+00:00 | 49.9 | 37.4% | 59.76 | 0.87s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:24:53.430831+00:00 | 70.4 | 85.1% | 102.12 | 0.56s |
#10 · Cross-device average · llama.cpp · BASELINE
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
56.4Tater score
60.9% accuracy63.28 tok/s0.57s TTFT10.2 GiB peak RSS2 devices4 runs
Chat81.25%Routing48.19%Spudex45.00%Synthesis90.00%Tool Accuracy69.14%
See 4 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:03:48.958884+00:00 | 73.3 | 91.8% | 83.12 | 0.51s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:42:29.439802+00:00 | 39.5 | 30.1% | 43.24 | 0.63s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:16:21.092711+00:00 | 39.5 | 30.1% | 43.21 | 0.63s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:22:56.307368+00:00 | 73.4 | 91.8% | 83.53 | 0.51s |
#1 · Device average · llama.cpp · MTP
TaterTotterson/Qwen3.6-35B-A3B-MTP-GGUF-Tater-NoThink
Qwen3.6-35B-A3B-UD-Q4_K_M.gguf
89.1Tater score
88.8% accuracy78.66 tok/s0.85s TTFT1.6 GiB peak RSS1 device2 runs
Chat100.00%Routing89.88%Spudex78.33%Synthesis100.00%Tool Accuracy87.43%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:48:25.786706+00:00 | 89.0 | 88.8% | 78.42 | 0.85s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:22:15.434255+00:00 | 89.1 | 88.8% | 78.90 | 0.85s |
#2 · Device average · llama.cpp · MTP
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
88.3Tater score
89.9% accuracy73.31 tok/s0.67s TTFT2.6 GiB peak RSS1 device2 runs
Chat100.00%Routing89.88%Spudex90.00%Synthesis90.00%Tool Accuracy88.14%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:43:16.924204+00:00 | 88.3 | 89.9% | 73.35 | 0.67s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:17:07.547513+00:00 | 88.3 | 89.9% | 73.27 | 0.67s |
#3 · Device average · llama.cpp · DFLASH
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
79.4Tater score
89.9% accuracy43.25 tok/s0.72s TTFT4.5 GiB peak RSS1 device2 runs
Chat100.00%Routing89.88%Spudex90.00%Synthesis90.00%Tool Accuracy88.14%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:44:15.074009+00:00 | 79.4 | 89.9% | 43.13 | 0.72s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:18:05.387711+00:00 | 79.5 | 89.9% | 43.37 | 0.72s |
#4 · Device average · llama.cpp · MTP
TaterTotterson/Qwen3.8-27B-GGUF-Tater-NoThink
Qwen3.8-27B-Tater-NoThink-Q4_K_M.gguf
70.2Tater score
88.8% accuracy21.19 tok/s2.35s TTFT4.0 GiB peak RSS1 device2 runs
Chat100.00%Routing85.29%Spudex90.00%Synthesis100.00%Tool Accuracy87.43%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:55:28.279842+00:00 | 70.3 | 88.8% | 21.20 | 2.35s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:29:18.938662+00:00 | 70.2 | 88.8% | 21.17 | 2.34s |
#5 · Device average · llama.cpp · BASELINE
TaterTotterson/Qwen3.6-35B-A3B-MTP-GGUF-Tater-NoThink
Qwen3.6-35B-A3B-UD-Q4_K_M.gguf
65.9Tater score
64.7% accuracy52.53 tok/s0.82s TTFT1.6 GiB peak RSS1 device2 runs
Chat100.00%Routing66.54%Spudex0.00%Synthesis35.00%Tool Accuracy87.43%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:47:02.097192+00:00 | 65.9 | 64.7% | 52.47 | 0.82s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:20:51.859924+00:00 | 65.9 | 64.7% | 52.58 | 0.82s |
#6 · Device average · llama.cpp · BASELINE
TaterTotterson/Qwen3.8-27B-GGUF-Tater-NoThink
Qwen3.8-27B-Tater-NoThink-Q4_K_M.gguf
57.5Tater score
73.2% accuracy10.97 tok/s2.32s TTFT2.5 GiB peak RSS1 device2 runs
Chat100.00%Routing82.27%Spudex0.00%Synthesis90.00%Tool Accuracy84.57%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:52:11.229193+00:00 | 57.5 | 73.1% | 10.96 | 2.32s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:26:01.254512+00:00 | 57.6 | 73.3% | 10.97 | 2.32s |
#7 · Device average · llama.cpp · BASELINE
TaterTotterson/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF-Tater-NoThink
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Tater-NoThink-Q4_0.gguf
49.9Tater score
37.4% accuracy59.51 tok/s0.87s TTFT1.1 GiB peak RSS1 device2 runs
Chat72.50%Routing0.00%Spudex0.00%Synthesis90.00%Tool Accuracy68.86%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:45:37.242320+00:00 | 49.8 | 37.4% | 59.27 | 0.87s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:19:27.390900+00:00 | 49.9 | 37.4% | 59.76 | 0.87s |
#8 · Device average · llama.cpp · BASELINE
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
39.5Tater score
30.1% accuracy43.22 tok/s0.63s TTFT2.1 GiB peak RSS1 device2 runs
Chat62.50%Routing6.50%Spudex0.00%Synthesis90.00%Tool Accuracy45.14%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-20T00:42:29.439802+00:00 | 39.5 | 30.1% | 43.24 | 0.63s |
| AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · 64.0 GiB | 2026-08-15T12:16:21.092711+00:00 | 39.5 | 30.1% | 43.21 | 0.63s |
#1 · Device average · llama.cpp · MTP
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
75.7Tater score
91.8% accuracy108.68 tok/s0.52s TTFT19.0 GiB peak RSS1 device2 runs
Chat100.00%Routing89.88%Spudex90.00%Synthesis90.00%Tool Accuracy93.14%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:04:13.758873+00:00 | 75.7 | 91.8% | 108.61 | 0.52s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:23:20.647032+00:00 | 75.7 | 91.8% | 108.76 | 0.52s |
#2 · Device average · mlx · BASELINE
TaterTotterson/Gemma-4-26B-A4B-IT-UD-Q4_K_XL-mlx-Tater-NoThink
674e8a504329b470c761c2e1feeb31d93d6ceeb7
73.6Tater score
95.3% accuracy49.77 tok/s0.26s TTFT18.5 GiB peak RSS1 device2 runs
Chat90.00%Routing91.88%Spudex96.67%Synthesis90.00%Tool Accuracy100.00%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:14:59.452371+00:00 | 73.6 | 95.3% | 49.54 | 0.26s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:33:45.435853+00:00 | 73.6 | 95.3% | 50.01 | 0.27s |
#3 · Device average · llama.cpp · BASELINE
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
73.3Tater score
91.8% accuracy83.33 tok/s0.51s TTFT18.2 GiB peak RSS1 device2 runs
Chat100.00%Routing89.88%Spudex90.00%Synthesis90.00%Tool Accuracy93.14%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:03:48.958884+00:00 | 73.3 | 91.8% | 83.12 | 0.51s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:22:56.307368+00:00 | 73.4 | 91.8% | 83.53 | 0.51s |
#4 · Device average · llama.cpp · MTP
TaterTotterson/Qwen3.6-35B-A3B-MTP-GGUF-Tater-NoThink
Qwen3.6-35B-A3B-UD-Q4_K_M.gguf
71.3Tater score
87.0% accuracy98.94 tok/s0.55s TTFT23.2 GiB peak RSS1 device2 runs
Chat100.00%Routing79.88%Spudex90.00%Synthesis100.00%Tool Accuracy87.43%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:07:29.363767+00:00 | 71.1 | 87.0% | 96.80 | 0.55s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:26:32.994760+00:00 | 71.5 | 87.0% | 101.07 | 0.55s |
#5 · Device average · llama.cpp · BASELINE
TaterTotterson/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF-Tater-NoThink
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Tater-NoThink-Q4_0.gguf
70.3Tater score
85.1% accuracy101.10 tok/s0.56s TTFT17.4 GiB peak RSS1 device2 runs
Chat100.00%Routing81.54%Spudex78.33%Synthesis90.00%Tool Accuracy87.43%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:05:48.227024+00:00 | 70.2 | 85.1% | 100.08 | 0.57s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:24:53.430831+00:00 | 70.4 | 85.1% | 102.12 | 0.56s |
#6 · Device average · llama.cpp · BASELINE
TaterTotterson/Qwen3.6-35B-A3B-MTP-GGUF-Tater-NoThink
Qwen3.6-35B-A3B-UD-Q4_K_M.gguf
69.5Tater score
87.0% accuracy78.77 tok/s0.53s TTFT20.8 GiB peak RSS1 device2 runs
Chat100.00%Routing79.88%Spudex90.00%Synthesis100.00%Tool Accuracy87.43%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:06:44.120075+00:00 | 69.4 | 87.0% | 78.22 | 0.53s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:25:48.331279+00:00 | 69.5 | 87.0% | 79.31 | 0.53s |
#7 · Device average · llama.cpp · DFLASH
TaterTotterson/gemma-4-26B-A4B-it-GGUF-Tater-NoThink
gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
68.9Tater score
91.8% accuracy36.30 tok/s0.53s TTFT21.4 GiB peak RSS1 device2 runs
Chat100.00%Routing89.88%Spudex90.00%Synthesis90.00%Tool Accuracy93.14%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:04:56.554299+00:00 | 68.9 | 91.8% | 36.20 | 0.52s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:24:03.026657+00:00 | 68.9 | 91.8% | 36.41 | 0.53s |
#8 · Device average · llama.cpp · BASELINE
unsloth/Qwen3.5-0.8B-GGUF
Qwen3.5-0.8B-Q4_K_M.gguf
67.2Tater score
53.1% accuracy214.21 tok/s0.10s TTFT1.2 GiB peak RSS1 device1 run
Chat100.00%Routing35.12%Spudex36.67%Synthesis90.83%Tool Accuracy60.05%
See 1 individual run
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:14:21.901105+00:00 | 67.2 | 53.1% | 214.21 | 0.10s |
#9 · Device average · llama.cpp · BASELINE
TaterTotterson/Qwen3.8-27B-GGUF-Tater-NoThink
Qwen3.8-27B-Tater-NoThink-Q4_K_M.gguf
65.2Tater score
88.8% accuracy26.85 tok/s2.76s TTFT18.8 GiB peak RSS1 device2 runs
Chat100.00%Routing85.29%Spudex90.00%Synthesis100.00%Tool Accuracy87.43%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:10:50.330144+00:00 | 65.2 | 88.8% | 26.61 | 2.78s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:29:50.869875+00:00 | 65.2 | 88.8% | 27.08 | 2.73s |
#10 · Device average · llama.cpp · MTP
TaterTotterson/Qwen3.8-27B-GGUF-Tater-NoThink
Qwen3.8-27B-Tater-NoThink-Q4_K_M.gguf
64.9Tater score
88.8% accuracy24.46 tok/s2.81s TTFT24.0 GiB peak RSS1 device2 runs
Chat100.00%Routing85.29%Spudex90.00%Synthesis100.00%Tool Accuracy87.43%
See 2 individual runs
| Device | Finished | Score | Accuracy | tok/s | TTFT |
|---|
| Apple M3 Ultra · 96.0 GiB | 2026-08-20T01:14:11.796656+00:00 | 64.9 | 88.8% | 24.75 | 2.82s |
| Apple M3 Ultra · 96.0 GiB | 2026-08-14T21:33:10.300966+00:00 | 64.9 | 88.8% | 24.18 | 2.79s |
