ERNIE-Light-DT-1.0 (high) vs Inkling 2 Mini (local Q4) (minimal)
The average score is effectively tied at 8.6 vs 8.7. Inkling 2 Mini (local Q4) (minimal) has the lower benchmark cost at ~$0.206 vs $1.027. ERNIE-Light-DT-1.0 (high) is faster at 25.66s vs 108.28s, with pass rates of 75.4% vs 73.9%.
Last updated at:
2026-11-22
Compared models
Rank
#113
Total Output Tokens
194,474
Response Time (avg)
25.66s
Total Cost
$1.027
Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
Rank
#108
Total Output Tokens
219,084
Response Time (avg)
108.28s
Total Cost
~$0.206Estimated GPU electricity cost only. It excludes the rest of the computer and hardware ownership. NVIDIA GeForce RTX 3090: 2.0753 h × 0.300 kW × €0.2896/kWh × 1.138 USD/EUR = ~$0.205185.
Recommended modelInkling 2 Mini (local Q4) (minimal)
It has the best score here (8.7), while costing about 5.0x less than ERNIE-Light-DT-1.0 (high).
Inkling 2 Mini (local Q4)Inkling 2 Mini (local Q4)minimalCandidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.Release: 2026-11-20
Inkling 2 Mini (local Q4)Inkling 2 Mini (local Q4)minimalCandidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.Release: 2026-11-20
Score
8.6Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
8.7Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
Rank
#113
#108
Reliability
9.3First-attempt success score: 10.0 means no retryable target API or rate-limit failures before successful calls; tracked failures lower the score.…
9.8First-attempt success score: 10.0 means no retryable target API or rate-limit failures before successful calls; tracked failures lower the score.…
Consistency
9.3Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
Attempts
69/69
69/69
Tests Correct
A test is fully passed only if every run passed for that test.Wrong answer: 5Extra formatting: 1No answer: 1Response Time (avg)25.66sResponse Time (max)200.80sResponse Time (total)590.18sA test is fully passed only if every run passed for that test.…
A test is fully passed only if every run passed for that test.Wrong answer: 4Did not follow instructions: 1Invalid tool call: 1Response Time (avg)108.28sResponse Time (max)1322.88sResponse Time (total)2490.37sA test is fully passed only if every run passed for that test.…
Attempt pass rate
75.4%Attempt pass rate = passed attempts / total attempts across runs.…
73.9%Attempt pass rate = passed attempts / total attempts across runs.…
Flaky tests
2Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
Total Runs
69Total Runs…
69Total Runs…
Cost per result
6.418
~1.207Estimated GPU electricity cost only. It excludes the rest of the computer and hardware ownership. NVIDIA GeForce RTX 3090: 2.0753 h × 0.300 kW × €0.2896/kWh × 1.138 USD/EUR = ~$0.205185.
Total Cost
$1.027
~$0.206Estimated GPU electricity cost only. It excludes the rest of the computer and hardware ownership. NVIDIA GeForce RTX 3090: 2.0753 h × 0.300 kW × €0.2896/kWh × 1.138 USD/EUR = ~$0.205185.
Input Price
$1.130 / 1MInput Price…
N/AInput Price…
Output Price
$3.941 / 1MOutput Price…
N/AOutput Price…
Cache Read Price
$0.113 / 1MCache Read Price…
N/ACache Read Price…
Cache Write Price
N/ACache Write Price…
N/ACache Write Price…
Total Input Tokens
230,563Total Input Tokens…
309,808Total Input Tokens…
Output Tokens
6,576Output Tokens…
1,579Output Tokens…
Reasoning Tokens
187,898Reasoning Tokens…
217,505Reasoning Tokens…
Response Time (avg)
25.66sResponse Time (avg)…
108.28sResponse Time (avg)…
Response Time (max)
200.80sResponse Time (max)…
1322.88sResponse Time (max)…
Response Time (total)
590.18sResponse Time (total)…
2490.37sResponse Time (total)…
Parameters
~320B total (~18B active)
26.4B total (3.4B active)
Availability
Closed
Open source
Cache prices apply to input tokens. Reads reuse cached prompts; writes store them and can cost extra. Output tokens use the output price.
Model generation showcase
Hamster playing table tennis
Prompt: Create a detailed SVG illustration of a hamster playing table tennis.
#113 ERNIE-Light-DT-1.0
high
Cost
$0.067
Time
129.8s
Tokens
16,960 tok
#108 Thinking Machines: Inkling 2 Mini (local Q4)
minimalCandidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
Cost
~$0.002Estimated GPU electricity cost only. It excludes the rest of the computer and hardware ownership. NVIDIA GeForce RTX 3090: 0.0188 h × 0.300 kW × €0.2896/kWh × 1.138 USD/EUR = ~$0.001862.
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)52.84sResponse Time (max)52.84sResponse Time (total)52.84sA test is fully passed only if every run passed for that test.…
52.84sResponse Time (avg)…
128,875Total Input Tokens…
1,210Output Tokens…
3,884Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)59.04sResponse Time (max)59.04sResponse Time (total)59.04sA test is fully passed only if every run passed for that test.…
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)1.89sResponse Time (max)2.10sResponse Time (total)7.55sA test is fully passed only if every run passed for that test.…
1.89sResponse Time (avg)…
639Total Input Tokens…
266Output Tokens…
486Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)6.20sResponse Time (max)10.50sResponse Time (total)24.78sA test is fully passed only if every run passed for that test.…
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)14.96sResponse Time (max)23.42sResponse Time (total)44.89sA test is fully passed only if every run passed for that test.…
14.96sResponse Time (avg)…
7,317Total Input Tokens…
419Output Tokens…
16,035Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)225.36sResponse Time (max)477.83sResponse Time (total)676.09sA test is fully passed only if every run passed for that test.…
6.5Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
50.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Wrong answer: 1Response Time (avg)27.93sResponse Time (max)48.23sResponse Time (total)55.85sA test is fully passed only if every run passed for that test.…
27.93sResponse Time (avg)…
76,694Total Input Tokens…
3,454Output Tokens…
16,156Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
6.4Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
50.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Invalid tool call: 1Response Time (avg)671.08sResponse Time (max)1322.88sResponse Time (total)1342.15sA test is fully passed only if every run passed for that test.…
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)1.72sResponse Time (max)1.80sResponse Time (total)3.43sA test is fully passed only if every run passed for that test.…
1.72sResponse Time (avg)…
7,149Total Input Tokens…
210Output Tokens…
246Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)28.12sResponse Time (max)35.26sResponse Time (total)56.25sA test is fully passed only if every run passed for that test.…
3.6Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
7.2Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
22.2%Attempt pass rate = passed attempts / total attempts across runs.…
1Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Extra formatting: 1No answer: 1Wrong answer: 1Response Time (avg)134.82sResponse Time (max)200.80sResponse Time (total)404.47sA test is fully passed only if every run passed for that test.…
134.82sResponse Time (avg)…
759Total Input Tokens…
220Output Tokens…
149,271Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
7.7Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
66.7%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Wrong answer: 1Response Time (avg)88.16sResponse Time (max)147.25sResponse Time (total)264.48sA test is fully passed only if every run passed for that test.…
6.5Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
3.4Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
66.7%Attempt pass rate = passed attempts / total attempts across runs.…
1Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Wrong answer: 1Response Time (avg)2.02sResponse Time (max)2.02sResponse Time (total)2.02sA test is fully passed only if every run passed for that test.…
2.02sResponse Time (avg)…
498Total Input Tokens…
185Output Tokens…
309Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
5.3Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
0.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Wrong answer: 1Response Time (avg)6.44sResponse Time (max)6.44sResponse Time (total)6.44sA test is fully passed only if every run passed for that test.…
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)2.04sResponse Time (max)2.05sResponse Time (total)4.08sA test is fully passed only if every run passed for that test.…
2.04sResponse Time (avg)…
678Total Input Tokens…
72Output Tokens…
166Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
9.8Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)5.25sResponse Time (max)5.70sResponse Time (total)10.49sA test is fully passed only if every run passed for that test.…
7.7Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
66.7%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Wrong answer: 1Response Time (avg)2.16sResponse Time (max)3.08sResponse Time (total)6.49sA test is fully passed only if every run passed for that test.…
2.16sResponse Time (avg)…
672Total Input Tokens…
291Output Tokens…
311Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
6.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
33.3%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Did not follow instructions: 1Wrong answer: 1Response Time (avg)7.32sResponse Time (max)7.68sResponse Time (total)21.96sA test is fully passed only if every run passed for that test.…
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)4.33sResponse Time (max)4.33sResponse Time (total)4.33sA test is fully passed only if every run passed for that test.…
4.33sResponse Time (avg)…
7,075Total Input Tokens…
231Output Tokens…
149Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
10.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
100.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.No failed answers.Response Time (avg)7.52sResponse Time (max)7.52sResponse Time (total)7.52sA test is fully passed only if every run passed for that test.…
3.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
0.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Wrong answer: 1Response Time (avg)4.23sResponse Time (max)4.23sResponse Time (total)4.23sA test is fully passed only if every run passed for that test.…
4.23sResponse Time (avg)…
207Total Input Tokens…
18Output Tokens…
885Reasoning Tokens…
Inkling 2 Mini (local Q4)Candidate responses were generated on local hardware. OpenRouter was used only for judging. inkling-2-mini-ud-q4_k_xl · UD-Q4_K_XL · llama.cpp b12631-e300919cc · NVIDIA GeForce RTX 3090.
3.0Summarizes broad quality across our full private benchmark suite, so ranking reflects consistent performance.…
10.0Consistency score reflects run-to-run stability (10 = very consistent, even if consistently wrong).…
0.0%Attempt pass rate = passed attempts / total attempts across runs.…
0Flaky tests had mixed outcomes across runs (at least one pass and one fail).…
A test is fully passed only if every run passed for that test.Wrong answer: 1Response Time (avg)21.18sResponse Time (max)21.18sResponse Time (total)21.18sA test is fully passed only if every run passed for that test.…