Trillion Labs · Model profile
Tri-21B-think Preview
Intelligence, Performance & Price Analysis
Source: Canonical slug: tri-21b-think-preview · Canonical model registry at build time
Routing verdict
Capable everyday model
Good for routine tasks; route complex reasoning and premium workloads to stronger models.
Tri-21B-think Preview summary metrics
- Intelligence
- 14 scoreTri-21B-think Preview scores 14 (estimated) on the Artificial Analysis Intelligence Index, placing it above average among other open weight models of similar size (median: 9).
- Input Price
- Not published
- Output Price
- Not published
Assessment
Routing verdict and task fit
| Workload | Rating | Notes |
|---|---|---|
| Complex reasoning & agentic workflows | Viable | Intelligence score 14 handles routine reasoning but may struggle with open-ended agentic tasks. |
Evidence
Evidence charts
AA-Omniscience Index
AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. Scores range from -100 to 100, where 0 means as many correct as incorrect answers, and negative scores mean more incorrect than correct. · Evaluation results measured independently by Artificial Analysis
- Claude Fable 5 (with fallback)40.1
- Gemini 3.5 Flash22.7
- GPT-5.5 (xhigh)20.1
- Grok 4.3 (high)18.3
- Kimi K2.66.4
- Muse Spark4.1
- GLM-5.2 (max)4
- MiniMax-M31.4
- Nemotron 3 Ultra-0.833
- DeepSeek V4 Pro (max)-10
- Qwen3.6 27B-19.8
- Qwen3.6 35B A3B-21.4
- Gemma 4 31B-45.4
- Gemma 4 26B A4B-48.1
- gpt-oss-120b (high)-50
- Mistral Small 3.1-50.3
- Qwen3 32B-50.5
- NVIDIA Nemotron 3 Nano-51.6
- Qwen3.5 9B-52.5
- Qwen3.6 27B-52.8
| Model | AA-Omniscience Index |
|---|---|
| Claude Fable 5 (with fallback) | 40.1 |
| Gemini 3.5 Flash | 22.7 |
| GPT-5.5 (xhigh) | 20.1 |
| Grok 4.3 (high) | 18.3 |
| Kimi K2.6 | 6.4 |
| Muse Spark | 4.1 |
| GLM-5.2 (max) | 4 |
| MiniMax-M3 | 1.4 |
| Nemotron 3 Ultra | -0.833 |
| DeepSeek V4 Pro (max) | -10 |
| Qwen3.6 27B | -19.8 |
| Qwen3.6 35B A3B | -21.4 |
| Gemma 4 31B | -45.4 |
| Gemma 4 26B A4B | -48.1 |
| gpt-oss-120b (high) | -50 |
| Mistral Small 3.1 | -50.3 |
| Qwen3 32B | -50.5 |
| NVIDIA Nemotron 3 Nano | -51.6 |
| Qwen3.5 9B | -52.5 |
| Qwen3.6 27B | -52.8 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Chart interpretation
- Each bar is one model measured the same way, so this model can be read against the others. The chart lists its source and retrieval date.
Artificial Analysis Intelligence Index by Open Weights / Proprietary
Artificial Analysis Intelligence Index v4.1 incorporates 9 evaluations: GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR · Evaluation results measured independently by Artificial Analysis
- Claude Fable 5 (with fallback)59.9
- GPT-5.5 (xhigh)54.8
- GLM-5.2 (max)51.1
- Gemini 3.5 Flash50.2
- MiniMax-M344.4
- DeepSeek V4 Pro (max)44.3
- Kimi K2.644.2
- Muse Spark43.1
- Nemotron 3 Ultra37.8
- Grok 4.3 (high)37.6
- Qwen3.6 27B37.1
- Qwen3.6 35B A3B31.6
- Qwen3.6 27B30.5
- Gemma 4 31B29.4
- Gemma 4 26B A4B25.7
- gpt-oss-120b (high)23.8
- Qwen3.5 9B21.4
- Devstral Small 217.4
- gpt-oss-20b (high)14.9
- Mistral Small 3.114.7
| Model | Artificial Analysis Intelligence Index by Open Weights / Proprietary |
|---|---|
| Claude Fable 5 (with fallback) | 59.9 |
| GPT-5.5 (xhigh) | 54.8 |
| GLM-5.2 (max) | 51.1 |
| Gemini 3.5 Flash | 50.2 |
| MiniMax-M3 | 44.4 |
| DeepSeek V4 Pro (max) | 44.3 |
| Kimi K2.6 | 44.2 |
| Muse Spark | 43.1 |
| Nemotron 3 Ultra | 37.8 |
| Grok 4.3 (high) | 37.6 |
| Qwen3.6 27B | 37.1 |
| Qwen3.6 35B A3B | 31.6 |
| Qwen3.6 27B | 30.5 |
| Gemma 4 31B | 29.4 |
| Gemma 4 26B A4B | 25.7 |
| gpt-oss-120b (high) | 23.8 |
| Qwen3.5 9B | 21.4 |
| Devstral Small 2 | 17.4 |
| gpt-oss-20b (high) | 14.9 |
| Mistral Small 3.1 | 14.7 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Benchmark interpretation
- Higher values indicate stronger overall capability. Bars show how models compare on composite intelligence metrics.
Artificial Analysis Intelligence Index
Artificial Analysis Intelligence Index v4.1 incorporates 9 evaluations: GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR · Evaluation results measured independently by Artificial Analysis
- Claude Fable 5 (with fallback)59.9
- GPT-5.5 (xhigh)54.8
- GLM-5.2 (max)51.1
- Gemini 3.5 Flash50.2
- MiniMax-M344.4
- DeepSeek V4 Pro (max)44.3
- Kimi K2.644.2
- Muse Spark43.1
- Nemotron 3 Ultra37.8
- Grok 4.3 (high)37.6
- Qwen3.6 27B37.1
- Qwen3.6 35B A3B31.6
- Qwen3.6 27B30.5
- Gemma 4 31B29.4
- Gemma 4 26B A4B25.7
- gpt-oss-120b (high)23.8
- Qwen3.5 9B21.4
- Devstral Small 217.4
- gpt-oss-20b (high)14.9
- Mistral Small 3.114.7
| Model | Artificial Analysis Intelligence Index |
|---|---|
| Claude Fable 5 (with fallback) | 59.9 |
| GPT-5.5 (xhigh) | 54.8 |
| GLM-5.2 (max) | 51.1 |
| Gemini 3.5 Flash | 50.2 |
| MiniMax-M3 | 44.4 |
| DeepSeek V4 Pro (max) | 44.3 |
| Kimi K2.6 | 44.2 |
| Muse Spark | 43.1 |
| Nemotron 3 Ultra | 37.8 |
| Grok 4.3 (high) | 37.6 |
| Qwen3.6 27B | 37.1 |
| Qwen3.6 35B A3B | 31.6 |
| Qwen3.6 27B | 30.5 |
| Gemma 4 31B | 29.4 |
| Gemma 4 26B A4B | 25.7 |
| gpt-oss-120b (high) | 23.8 |
| Qwen3.5 9B | 21.4 |
| Devstral Small 2 | 17.4 |
| gpt-oss-20b (high) | 14.9 |
| Mistral Small 3.1 | 14.7 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Benchmark interpretation
- Higher values indicate stronger overall capability. Bars show how models compare on composite intelligence metrics.
Artificial Analysis Openness Index: Score
Openness Index assesses model openness on a 0 to 100 normalized scale (higher is more open) · Evaluation results measured independently by Artificial Analysis
- Nemotron 3 Ultra83.3
- NVIDIA Nemotron 3 Nano83.3
- DeepSeek V4 Pro (max)50
- Magistral Small 1.250
- GLM-5.2 (max)44.4
- Qwen3 30B A3B 250744.4
- gpt-oss-120b (high)38.9
- Qwen3.6 27B38.9
- Qwen3.6 35B A3B38.9
- Qwen3.6 27B38.9
- Gemma 4 31B38.9
- Gemma 4 26B A4B38.9
- Qwen3.5 9B38.9
- gpt-oss-20b (high)38.9
- Kimi K2.633.3
- MiniMax-M333.3
| Model | Artificial Analysis Openness Index: Score |
|---|---|
| Nemotron 3 Ultra | 83.3 |
| NVIDIA Nemotron 3 Nano | 83.3 |
| DeepSeek V4 Pro (max) | 50 |
| Magistral Small 1.2 | 50 |
| GLM-5.2 (max) | 44.4 |
| Qwen3 30B A3B 2507 | 44.4 |
| gpt-oss-120b (high) | 38.9 |
| Qwen3.6 27B | 38.9 |
| Qwen3.6 35B A3B | 38.9 |
| Qwen3.6 27B | 38.9 |
| Gemma 4 31B | 38.9 |
| Gemma 4 26B A4B | 38.9 |
| Qwen3.5 9B | 38.9 |
| gpt-oss-20b (high) | 38.9 |
| Kimi K2.6 | 33.3 |
| MiniMax-M3 | 33.3 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Chart interpretation
- Each bar is one model measured the same way, so this model can be read against the others. The chart lists its source and retrieval date.
Intelligence
Artificial Analysis Intelligence Index · Higher is better · Evaluation results measured independently by Artificial Analysis
- Claude Fable 5 (with fallback)59.9
- GPT-5.5 (xhigh)54.8
- GLM-5.2 (max)51.1
- Gemini 3.5 Flash50.2
- MiniMax-M344.4
- DeepSeek V4 Pro (max)44.3
- Kimi K2.644.2
- Muse Spark43.1
- Nemotron 3 Ultra37.8
- Grok 4.3 (high)37.6
- gpt-oss-120b (high)23.8
- Tri-21B-think Preview (this model)13.6
| Model | Intelligence |
|---|---|
| Claude Fable 5 (with fallback) | 59.9 |
| GPT-5.5 (xhigh) | 54.8 |
| GLM-5.2 (max) | 51.1 |
| Gemini 3.5 Flash | 50.2 |
| MiniMax-M3 | 44.4 |
| DeepSeek V4 Pro (max) | 44.3 |
| Kimi K2.6 | 44.2 |
| Muse Spark | 43.1 |
| Nemotron 3 Ultra | 37.8 |
| Grok 4.3 (high) | 37.6 |
| gpt-oss-120b (high) | 23.8 |
| Tri-21B-think Preview (this model) | 13.6 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Benchmark interpretation
- Higher values indicate stronger overall capability. Bars show how models compare on composite intelligence metrics.
Speed
Output tokens per second · Higher is better · Evaluation results measured independently by Artificial Analysis
- gpt-oss-120b (high)314 t/s
- Nemotron 3 Ultra250 t/s
- GLM-5.2 (max)218 t/s
- Gemini 3.5 Flash192 t/s
- Grok 4.3 (high)164 t/s
- MiniMax-M399 t/s
- GPT-5.5 (xhigh)88 t/s
- Kimi K2.676 t/s
- DeepSeek V4 Pro (max)72 t/s
- Claude Fable 5 (with fallback)71 t/s
| Model | Speed |
|---|---|
| gpt-oss-120b (high) | 314 t/s |
| Nemotron 3 Ultra | 250 t/s |
| GLM-5.2 (max) | 218 t/s |
| Gemini 3.5 Flash | 192 t/s |
| Grok 4.3 (high) | 164 t/s |
| MiniMax-M3 | 99 t/s |
| GPT-5.5 (xhigh) | 88 t/s |
| Kimi K2.6 | 76 t/s |
| DeepSeek V4 Pro (max) | 72 t/s |
| Claude Fable 5 (with fallback) | 71 t/s |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Output speed context
- Speed should be read together with latency and cost; fast output does not always mean better user-perceived performance.
Context Window
Context window: tokens limit · Higher is better · Evaluation results measured independently by Artificial Analysis
- Gemini 3.5 Flash1M
- Claude Fable 5 (with fallback)1M
- GLM-5.2 (max)1M
- DeepSeek V4 Pro (max)1M
- Grok 4.3 (high)1M
- MiniMax-M31M
- NVIDIA Nemotron 3 Nano1M
- GPT-5.5 (xhigh)922k
- Nemotron 3 Ultra262.1k
- Muse Spark262.1k
- Qwen3.6 27B262.1k
- Qwen3.6 35B A3B262.1k
- Qwen3.6 27B262.1k
- Qwen3.5 9B262.1k
- Qwen3 30B A3B 2507262.1k
- Kimi K2.6256k
- Gemma 4 31B256k
- Gemma 4 26B A4B256k
- Devstral Small 2256k
- DiffusionGemma 26B A4B256k
| Model | Context Window |
|---|---|
| Gemini 3.5 Flash | 1M |
| Claude Fable 5 (with fallback) | 1M |
| GLM-5.2 (max) | 1M |
| DeepSeek V4 Pro (max) | 1M |
| Grok 4.3 (high) | 1M |
| MiniMax-M3 | 1M |
| NVIDIA Nemotron 3 Nano | 1M |
| GPT-5.5 (xhigh) | 922k |
| Nemotron 3 Ultra | 262.1k |
| Muse Spark | 262.1k |
| Qwen3.6 27B | 262.1k |
| Qwen3.6 35B A3B | 262.1k |
| Qwen3.6 27B | 262.1k |
| Qwen3.5 9B | 262.1k |
| Qwen3 30B A3B 2507 | 262.1k |
| Kimi K2.6 | 256k |
| Gemma 4 31B | 256k |
| Gemma 4 26B A4B | 256k |
| Devstral Small 2 | 256k |
| DiffusionGemma 26B A4B | 256k |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Context window sizing
- Larger context windows enable complex multi-turn tasks but may increase latency and cost for long inputs.
Cost per Intelligence Index Task
Weighted average cost (USD) per Artificial Analysis Intelligence Index task, segmented by token type. Lower is better · Evaluation results measured independently by Artificial Analysis
- gpt-oss-20b (high)0.000555
- NVIDIA Nemotron 3 Nano0.000622
- Gemma 4 26B A4B0.00194
- Mistral Small 3.10.00143
- DeepSeek V4 Pro (max)0.0063
- gpt-oss-120b (high)0.00454
- Qwen3 30B A3B 25070.00466
- MiniMax-M30.0139
- Grok 4.3 (high)0.0101
- Ministral 3 14B0.00235
- Qwen3.5 9B0.00178
- Qwen3.6 35B A3B0.0213
- Nemotron 3 Ultra0.0209
- Magistral Small 1.20.00704
- Qwen3.6 27B0.0251
- Kimi K2.60.0216
- Qwen3.6 27B0.041
- GLM-5.2 (max)0.0268
- Gemini 3.5 Flash0.0884
- GPT-5.5 (xhigh)0.107
| Model | Cost per Intelligence Index Task |
|---|---|
| gpt-oss-20b (high) | 0.000555 |
| NVIDIA Nemotron 3 Nano | 0.000622 |
| Gemma 4 26B A4B | 0.00194 |
| Mistral Small 3.1 | 0.00143 |
| DeepSeek V4 Pro (max) | 0.0063 |
| gpt-oss-120b (high) | 0.00454 |
| Qwen3 30B A3B 2507 | 0.00466 |
| MiniMax-M3 | 0.0139 |
| Grok 4.3 (high) | 0.0101 |
| Ministral 3 14B | 0.00235 |
| Qwen3.5 9B | 0.00178 |
| Qwen3.6 35B A3B | 0.0213 |
| Nemotron 3 Ultra | 0.0209 |
| Magistral Small 1.2 | 0.00704 |
| Qwen3.6 27B | 0.0251 |
| Kimi K2.6 | 0.0216 |
| Qwen3.6 27B | 0.041 |
| GLM-5.2 (max) | 0.0268 |
| Gemini 3.5 Flash | 0.0884 |
| GPT-5.5 (xhigh) | 0.107 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Benchmark interpretation
- Higher values indicate stronger overall capability. Bars show how models compare on composite intelligence metrics.
- Cost pressure
- Ethen Gateway should route away from this model when the task is simple, repetitive, or likely to generate many output tokens.
Cost per Task
Weighted average cost (USD) per Intelligence Index task · Lower is better · Evaluation results measured independently by Artificial Analysis
- DeepSeek V4 Pro (max)$0.045
- gpt-oss-120b (high)$0.061
- MiniMax-M3$0.125
- Grok 4.3 (high)$0.139
- Nemotron 3 Ultra$0.244
- Kimi K2.6$0.346
- GLM-5.2 (max)$0.370
- Gemini 3.5 Flash$0.586
- GPT-5.5 (xhigh)$0.863
- Claude Fable 5 (with fallback)$2.75
| Model | Cost per Task |
|---|---|
| DeepSeek V4 Pro (max) | $0.045 |
| gpt-oss-120b (high) | $0.061 |
| MiniMax-M3 | $0.125 |
| Grok 4.3 (high) | $0.139 |
| Nemotron 3 Ultra | $0.244 |
| Kimi K2.6 | $0.346 |
| GLM-5.2 (max) | $0.370 |
| Gemini 3.5 Flash | $0.586 |
| GPT-5.5 (xhigh) | $0.863 |
| Claude Fable 5 (with fallback) | $2.75 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Cost pressure
- Ethen Gateway should route away from this model when the task is simple, repetitive, or likely to generate many output tokens.
Cost to Run Artificial Analysis Intelligence Index
Cost (USD) to run all evaluations in the Artificial Analysis Intelligence Index · Evaluation results measured independently by Artificial Analysis
- gpt-oss-20b (high)0.795
- Mistral Small 3.11.5
- NVIDIA Nemotron 3 Nano0.658
- Gemma 4 26B A4B2.3
- gpt-oss-120b (high)3.5
- Ministral 3 14B3
- Qwen3 30B A3B 25077.6
- DeepSeek V4 Pro (max)7.2
- Qwen3.5 9B2.6
- MiniMax-M314.5
- Grok 4.3 (high)12.5
- Magistral Small 1.28.2
- Qwen3.6 35B A3B164.7
- Qwen3.6 27B87.9
- Nemotron 3 Ultra33.5
- Qwen3.6 27B28
- GLM-5.2 (max)32.9
- Kimi K2.625
- Gemini 3.5 Flash114
- GPT-5.5 (xhigh)131.8
| Model | Cost to Run Artificial Analysis Intelligence Index |
|---|---|
| gpt-oss-20b (high) | 0.795 |
| Mistral Small 3.1 | 1.5 |
| NVIDIA Nemotron 3 Nano | 0.658 |
| Gemma 4 26B A4B | 2.3 |
| gpt-oss-120b (high) | 3.5 |
| Ministral 3 14B | 3 |
| Qwen3 30B A3B 2507 | 7.6 |
| DeepSeek V4 Pro (max) | 7.2 |
| Qwen3.5 9B | 2.6 |
| MiniMax-M3 | 14.5 |
| Grok 4.3 (high) | 12.5 |
| Magistral Small 1.2 | 8.2 |
| Qwen3.6 35B A3B | 164.7 |
| Qwen3.6 27B | 87.9 |
| Nemotron 3 Ultra | 33.5 |
| Qwen3.6 27B | 28 |
| GLM-5.2 (max) | 32.9 |
| Kimi K2.6 | 25 |
| Gemini 3.5 Flash | 114 |
| GPT-5.5 (xhigh) | 131.8 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Benchmark interpretation
- Higher values indicate stronger overall capability. Bars show how models compare on composite intelligence metrics.
- Cost pressure
- Ethen Gateway should route away from this model when the task is simple, repetitive, or likely to generate many output tokens.
Pricing: Cache Hit, Input, and Output
Price (USD per M Tokens) · Evaluation results measured independently by Artificial Analysis
- Input
- Output
- Cache Hit
- Qwen3.5 9B$0.250
- gpt-oss-20b (high)$0.250
- NVIDIA Nemotron 3 Nano$0.250
- Mistral Small 3.1$0.400
- Ministral 3 14B$0.400
- Gemma 4 26B A4B$0.615
- gpt-oss-120b (high)$0.900
- DeepSeek V4 Pro (max)$1.31
- MiniMax-M3$1.56
- Qwen3.6 35B A3B$1.73
- Magistral Small 1.2$2.00
- Qwen3 30B A3B 2507$2.60
- Nemotron 3 Ultra$3.60
- Grok 4.3 (high)$3.95
- Qwen3.6 27B$4.20
- Qwen3.6 27B$4.20
- Kimi K2.6$5.11
- GLM-5.2 (max)$6.06
- Qwen3 32B$9.10
- Gemini 3.5 Flash$10.65
| Model | Input | Output | Cache Hit | Total |
|---|---|---|---|---|
| Qwen3.5 9B | $0.100 | $0.150 | Not published | $0.250 |
| gpt-oss-20b (high) | $0.050 | $0.200 | Not published | $0.250 |
| NVIDIA Nemotron 3 Nano | $0.050 | $0.200 | Not published | $0.250 |
| Mistral Small 3.1 | $0.100 | $0.300 | Not published | $0.400 |
| Ministral 3 14B | $0.200 | $0.200 | Not published | $0.400 |
| Gemma 4 26B A4B | $0.130 | $0.400 | $0.085 | $0.615 |
| gpt-oss-120b (high) | $0.150 | $0.600 | $0.150 | $0.900 |
| DeepSeek V4 Pro (max) | $0.435 | $0.870 | $0.0036 | $1.31 |
| MiniMax-M3 | $0.300 | $1.20 | $0.060 | $1.56 |
| Qwen3.6 35B A3B | $0.248 | $1.49 | Not published | $1.73 |
| Magistral Small 1.2 | $0.500 | $1.50 | Not published | $2.00 |
| Qwen3 30B A3B 2507 | $0.200 | $2.40 | Not published | $2.60 |
| Nemotron 3 Ultra | $0.675 | $2.67 | $0.250 | $3.60 |
| Grok 4.3 (high) | $1.25 | $2.50 | $0.200 | $3.95 |
| Qwen3.6 27B | $0.600 | $3.60 | Not published | $4.20 |
| Qwen3.6 27B | $0.600 | $3.60 | Not published | $4.20 |
| Kimi K2.6 | $0.950 | $4.00 | $0.160 | $5.11 |
| GLM-5.2 (max) | $1.40 | $4.40 | $0.260 | $6.06 |
| Qwen3 32B | $0.700 | $8.40 | Not published | $9.10 |
| Gemini 3.5 Flash | $1.50 | $9.00 | $0.150 | $10.65 |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Chart interpretation
- Each bar is one model measured the same way, so this model can be read against the others. The chart lists its source and retrieval date.
Model Size: Total and Active Parameters
Comparison between total model parameters and parameters active during inference · Evaluation results measured independently by Artificial Analysis
- Passive Params
- Active Params
- DeepSeek V4 Pro (max)1,600B
- Kimi K2.61,000B
- GLM-5.2 (max)753B
- Nemotron 3 Ultra550B
- MiniMax-M3428B
- gpt-oss-120b (high)117B
- Qwen3.6 35B A3B36B
- Qwen3 32B32.8B
- NVIDIA Nemotron 3 Nano31.6B
- Gemma 4 31B30.7B
- Qwen3 30B A3B 250730.5B
- Qwen3.6 27B27.8B
- Qwen3.6 27B27.8B
- Gemma 4 26B A4B25.2B
- DiffusionGemma 26B A4B25.2B
- Devstral Small 224B
- Mistral Small 3.124B
- Magistral Small 1.224B
- Tri-21B-think Preview (this model)21B
- gpt-oss-20b (high)21B
| Model | Passive Params | Active Params | Total |
|---|---|---|---|
| DeepSeek V4 Pro (max) | 1,551B | 49B | 1,600B |
| Kimi K2.6 | 968B | 32B | 1,000B |
| GLM-5.2 (max) | 713B | 40B | 753B |
| Nemotron 3 Ultra | 495B | 55B | 550B |
| MiniMax-M3 | 405B | 23B | 428B |
| gpt-oss-120b (high) | 111.9B | 5.1B | 117B |
| Qwen3.6 35B A3B | 33B | 3B | 36B |
| Qwen3 32B | 0B | 32.8B | 32.8B |
| NVIDIA Nemotron 3 Nano | 28B | 3.6B | 31.6B |
| Gemma 4 31B | 0B | 30.7B | 30.7B |
| Qwen3 30B A3B 2507 | 27.2B | 3.3B | 30.5B |
| Qwen3.6 27B | 0B | 27.8B | 27.8B |
| Qwen3.6 27B | 0B | 27.8B | 27.8B |
| Gemma 4 26B A4B | 21.4B | 3.8B | 25.2B |
| DiffusionGemma 26B A4B | 21.4B | 3.8B | 25.2B |
| Devstral Small 2 | 0B | 24B | 24B |
| Mistral Small 3.1 | 0B | 24B | 24B |
| Magistral Small 1.2 | 0B | 24B | 24B |
| Tri-21B-think Preview (this model) | 0B | 21B | 21B |
| gpt-oss-20b (high) | 17.4B | 3.6B | 21B |
Published benchmark dataset · Retrieved
Method and source
- Methodology
- Independent test run by Artificial Analysis on dedicated hardware.
- Architecture scale
- Larger models generally perform better but cost more to serve. Active parameter count matters for inference efficiency.
Specifications
Architecture and limits
| Specification | Value | Source |
|---|---|---|
| Model type | Open weights | faq |
| Reasoning | Yes | faq |
| Input modalities | Tri-21B-think Preview supports text input. | faq |
| Output modalities | Tri-21B-think Preview supports text output. | faq |
| Context window | 32k tokens | faq |
| Open weights / source | Yes, Tri-21B-think Preview is open weights. The model weights are publicly available and can be downloaded for self-hosting. | faq |
| Parameters | Tri-21B-think Preview has 21 billion parameters. | faq |
| License | Tri-21B-think Preview is released under the Apache 2.0 license. This license allows commercial use. | faq |
Method
Methodology & Provenance
This page is rendered from the normalized profile and page JSON for Tri-21B-think Preview.
Benchmark values are preserved as normalized; only layout, disclosure ordering, and typography are adjusted for readability.
Questions
Model FAQ and technical disclosures
Models
Related
- 01Trillion LabsOther Trillion Labs models in the catalog.
- 02LeaderboardsCertified metric ranks, when published.