Overview
Local language models compared: knowledge, math, code and German, plus speed. Each model variant is shown with its most recent completed run. How the numbers are measured is described under Methodology.
Suite v2b · profile quick
| Model | Overall | Knowledge & reasoning | Math | Instruction following | Code | German | gen t/s | gen t/s run | pp t/s 4k | TTFT 4k ms | Date |
|---|---|---|---|---|---|---|---|---|---|---|---|
|
gpt-oss-120b
MXFP4 · Reasoning on · direct · NVIDIA DGX Spark, 128 GB RAM
|
73.1 | 68.5 | 70.4 | 66.1 | 83.0 | 77.4 | – | 20 | – | – | 2026-09-18 |
|
qwen3.8-flash-next-abliterated (pool)
W4A16-NVFP4 · Reasoning off · direct · 2× NVIDIA DGX Spark, je 128 GB RAM
|
72.5 | 77.2 | 72.0 | 55.2 | 77.0 | 81.4 | 41 | 37 | 2569 | 1566 | 2026-09-27 |
|
qwen3.8-flash-next-ple4
W4A16 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
71.5 | 77.9 | 73.2 | 49.6 | 75.5 | 81.1 | 26 | 22 | 2041 | 1974 | 2026-09-17 |
|
gemma-4-12b-it
Q8 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
65.8 | 67.5 | 60.8 | 51.3 | 75.0 | 74.5 | 17 | 14 | 1359 | 2952 | 2026-09-18 |
|
gemma-4-12b-it-heretic
Q8 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
65.8 | 68.4 | 59.2 | 50.0 | 75.0 | 76.5 | 17 | 14 | 1358 | 2954 | 2026-09-19 |
|
qwen3-coder-next
Q6 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
65.1 | 71.3 | 61.6 | 45.9 | 70.5 | 76.0 | 44 | 24 | 988 | 4078 | 2026-09-23 |
|
qwen3.8-27b-mtp
NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
64.2 | 69.9 | 60.4 | 49.5 | 65.0 | 76.3 | 15 | 22 | 2551 | 1578 | 2026-09-18 |
|
occamy-1.0
NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
64.0 | 73.1 | 62.4 | 44.7 | 63.0 | 76.9 | 37 | 32 | 5539 | 726 | 2026-09-21 |
|
qwen3.8-27b-abliterated
NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
61.4 | 71.5 | 58.0 | 46.1 | 59.5 | 72.1 | 16 | 22 | 2472 | 1628 | 2026-09-20 |
|
ternary-bonsai-2-27b
FP8 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
61.3 | 69.5 | 59.6 | 46.3 | 56.5 | 74.5 | 28 | 19 | 898 | 4480 | 2026-09-21 |
|
mimo-v2.6-flash (pool)
FP8 + MXFP4 · Reasoning off · direct · 2× NVIDIA DGX Spark, je 128 GB RAM
|
59.4 | 69.7 | 41.6 | 49.8 | 61.5 | 74.5 | 27 | 22 | 2178 | 1848 | 2026-09-26 |
|
gemma-4-12b-it-heretic
NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
57.7 | 63.7 | 48.4 | 46.8 | 62.0 | 67.7 | 21 | 21 | 5258 | 763 | 2026-09-18 |
|
nemotron-3.5-lightning-dspark
NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
53.1 | 55.8 | 40.4 | 46.0 | 53.0 | 70.5 | 84 | 70 | 7237 | 554 | 2026-09-17 |
|
qwen3.5-9b-heretic-v2-mtp
NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
51.7 | 65.1 | 35.6 | 44.1 | 46.5 | 67.0 | 32 | 35 | 6291 | 640 | 2026-09-18 |
|
zgcm-1-7b
NVFP4 · Reasoning off · direct · NVIDIA DGX Spark, 128 GB RAM
|
33.4 | 34.7 | 40.0 | 44.3 | 23.0 | 24.8 | 34 | 34 | 3958 | 1023 | 2026-09-23 |
x-axis: generation per request measured during the run; hollow points only have the speed-test value.
Scores run from 0 to 100; the best value per category is shown in bold. Hovering a category cell lists its tasks with standard error (±) and number of items (n). Speed: generation in tokens/s from the speed test and measured during the run (gen t/s run, per request), prompt processing at 4k context in tokens/s, time to first token at 4k context in ms. Click a column header to sort.