Skip to content

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

Tri-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).
Task fit assessment
WorkloadRatingNotes
Complex reasoning & agentic workflowsViableIntelligence score 14 handles routine reasoning but may struggle with open-ended agentic tasks.

Evidence

Evidence charts

12 published comparison charts for this profile. Tri-21B-think Preview is marked with a diamond and bold label; each chart has a table view and lists its source and retrieval date.

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

20 models in the published order. Tri-21B-think Preview is not marked in this published set of 20 models.
  1. Claude Fable 5 (with fallback)40.1
  2. Gemini 3.5 Flash22.7
  3. GPT-5.5 (xhigh)20.1
  4. Grok 4.3 (high)18.3
  5. Kimi K2.66.4
  6. Muse Spark4.1
  7. GLM-5.2 (max)4
  8. MiniMax-M31.4
  9. Nemotron 3 Ultra-0.833
  10. DeepSeek V4 Pro (max)-10
  11. Qwen3.6 27B-19.8
  12. Qwen3.6 35B A3B-21.4
  13. Gemma 4 31B-45.4
  14. Gemma 4 26B A4B-48.1
  15. gpt-oss-120b (high)-50
  16. Mistral Small 3.1-50.3
  17. Qwen3 32B-50.5
  18. NVIDIA Nemotron 3 Nano-51.6
  19. Qwen3.5 9B-52.5
  20. 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

20 models in the published order. Tri-21B-think Preview is not marked in this published set of 20 models.
  1. Claude Fable 5 (with fallback)59.9
  2. GPT-5.5 (xhigh)54.8
  3. GLM-5.2 (max)51.1
  4. Gemini 3.5 Flash50.2
  5. MiniMax-M344.4
  6. DeepSeek V4 Pro (max)44.3
  7. Kimi K2.644.2
  8. Muse Spark43.1
  9. Nemotron 3 Ultra37.8
  10. Grok 4.3 (high)37.6
  11. Qwen3.6 27B37.1
  12. Qwen3.6 35B A3B31.6
  13. Qwen3.6 27B30.5
  14. Gemma 4 31B29.4
  15. Gemma 4 26B A4B25.7
  16. gpt-oss-120b (high)23.8
  17. Qwen3.5 9B21.4
  18. Devstral Small 217.4
  19. gpt-oss-20b (high)14.9
  20. Mistral Small 3.114.7

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

20 models in the published order. Tri-21B-think Preview is not marked in this published set of 20 models.
  1. Claude Fable 5 (with fallback)59.9
  2. GPT-5.5 (xhigh)54.8
  3. GLM-5.2 (max)51.1
  4. Gemini 3.5 Flash50.2
  5. MiniMax-M344.4
  6. DeepSeek V4 Pro (max)44.3
  7. Kimi K2.644.2
  8. Muse Spark43.1
  9. Nemotron 3 Ultra37.8
  10. Grok 4.3 (high)37.6
  11. Qwen3.6 27B37.1
  12. Qwen3.6 35B A3B31.6
  13. Qwen3.6 27B30.5
  14. Gemma 4 31B29.4
  15. Gemma 4 26B A4B25.7
  16. gpt-oss-120b (high)23.8
  17. Qwen3.5 9B21.4
  18. Devstral Small 217.4
  19. gpt-oss-20b (high)14.9
  20. Mistral Small 3.114.7

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

16 models in the published order. Tri-21B-think Preview is not marked in this published set of 16 models.
  1. Nemotron 3 Ultra83.3
  2. NVIDIA Nemotron 3 Nano83.3
  3. DeepSeek V4 Pro (max)50
  4. Magistral Small 1.250
  5. GLM-5.2 (max)44.4
  6. Qwen3 30B A3B 250744.4
  7. gpt-oss-120b (high)38.9
  8. Qwen3.6 27B38.9
  9. Qwen3.6 35B A3B38.9
  10. Qwen3.6 27B38.9
  11. Gemma 4 31B38.9
  12. Gemma 4 26B A4B38.9
  13. Qwen3.5 9B38.9
  14. gpt-oss-20b (high)38.9
  15. Kimi K2.633.3
  16. MiniMax-M333.3

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

12 models in the published order. Tri-21B-think Preview: 13.6, row 12 of 12 in the published order.
  1. Claude Fable 5 (with fallback)59.9
  2. GPT-5.5 (xhigh)54.8
  3. GLM-5.2 (max)51.1
  4. Gemini 3.5 Flash50.2
  5. MiniMax-M344.4
  6. DeepSeek V4 Pro (max)44.3
  7. Kimi K2.644.2
  8. Muse Spark43.1
  9. Nemotron 3 Ultra37.8
  10. Grok 4.3 (high)37.6
  11. gpt-oss-120b (high)23.8
  12. 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

10 models in the published order. Tri-21B-think Preview is not marked in this published set of 10 models.
  1. gpt-oss-120b (high)314 t/s
  2. Nemotron 3 Ultra250 t/s
  3. GLM-5.2 (max)218 t/s
  4. Gemini 3.5 Flash192 t/s
  5. Grok 4.3 (high)164 t/s
  6. MiniMax-M399 t/s
  7. GPT-5.5 (xhigh)88 t/s
  8. Kimi K2.676 t/s
  9. DeepSeek V4 Pro (max)72 t/s
  10. 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

20 models in the published order. Tri-21B-think Preview is not marked in this published set of 20 models.
  1. Gemini 3.5 Flash1M
  2. Claude Fable 5 (with fallback)1M
  3. GLM-5.2 (max)1M
  4. DeepSeek V4 Pro (max)1M
  5. Grok 4.3 (high)1M
  6. MiniMax-M31M
  7. NVIDIA Nemotron 3 Nano1M
  8. GPT-5.5 (xhigh)922k
  9. Nemotron 3 Ultra262.1k
  10. Muse Spark262.1k
  11. Qwen3.6 27B262.1k
  12. Qwen3.6 35B A3B262.1k
  13. Qwen3.6 27B262.1k
  14. Qwen3.5 9B262.1k
  15. Qwen3 30B A3B 2507262.1k
  16. Kimi K2.6256k
  17. Gemma 4 31B256k
  18. Gemma 4 26B A4B256k
  19. Devstral Small 2256k
  20. DiffusionGemma 26B A4B256k

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

20 models in the published order. Tri-21B-think Preview is not marked in this published set of 20 models.
  1. gpt-oss-20b (high)0.000555
  2. NVIDIA Nemotron 3 Nano0.000622
  3. Gemma 4 26B A4B0.00194
  4. Mistral Small 3.10.00143
  5. DeepSeek V4 Pro (max)0.0063
  6. gpt-oss-120b (high)0.00454
  7. Qwen3 30B A3B 25070.00466
  8. MiniMax-M30.0139
  9. Grok 4.3 (high)0.0101
  10. Ministral 3 14B0.00235
  11. Qwen3.5 9B0.00178
  12. Qwen3.6 35B A3B0.0213
  13. Nemotron 3 Ultra0.0209
  14. Magistral Small 1.20.00704
  15. Qwen3.6 27B0.0251
  16. Kimi K2.60.0216
  17. Qwen3.6 27B0.041
  18. GLM-5.2 (max)0.0268
  19. Gemini 3.5 Flash0.0884
  20. 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

10 models in the published order. Tri-21B-think Preview is not marked in this published set of 10 models.
  1. DeepSeek V4 Pro (max)$0.045
  2. gpt-oss-120b (high)$0.061
  3. MiniMax-M3$0.125
  4. Grok 4.3 (high)$0.139
  5. Nemotron 3 Ultra$0.244
  6. Kimi K2.6$0.346
  7. GLM-5.2 (max)$0.370
  8. Gemini 3.5 Flash$0.586
  9. GPT-5.5 (xhigh)$0.863
  10. 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

20 models in the published order. Tri-21B-think Preview is not marked in this published set of 20 models.
  1. gpt-oss-20b (high)0.795
  2. Mistral Small 3.11.5
  3. NVIDIA Nemotron 3 Nano0.658
  4. Gemma 4 26B A4B2.3
  5. gpt-oss-120b (high)3.5
  6. Ministral 3 14B3
  7. Qwen3 30B A3B 25077.6
  8. DeepSeek V4 Pro (max)7.2
  9. Qwen3.5 9B2.6
  10. MiniMax-M314.5
  11. Grok 4.3 (high)12.5
  12. Magistral Small 1.28.2
  13. Qwen3.6 35B A3B164.7
  14. Qwen3.6 27B87.9
  15. Nemotron 3 Ultra33.5
  16. Qwen3.6 27B28
  17. GLM-5.2 (max)32.9
  18. Kimi K2.625
  19. Gemini 3.5 Flash114
  20. 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

20 models; segments: Input, Output, Cache Hit. Tri-21B-think Preview is not marked in this published set of 20 models.
  • Input
  • Output
  • Cache Hit
  1. Qwen3.5 9B$0.250
  2. gpt-oss-20b (high)$0.250
  3. NVIDIA Nemotron 3 Nano$0.250
  4. Mistral Small 3.1$0.400
  5. Ministral 3 14B$0.400
  6. Gemma 4 26B A4B$0.615
  7. gpt-oss-120b (high)$0.900
  8. DeepSeek V4 Pro (max)$1.31
  9. MiniMax-M3$1.56
  10. Qwen3.6 35B A3B$1.73
  11. Magistral Small 1.2$2.00
  12. Qwen3 30B A3B 2507$2.60
  13. Nemotron 3 Ultra$3.60
  14. Grok 4.3 (high)$3.95
  15. Qwen3.6 27B$4.20
  16. Qwen3.6 27B$4.20
  17. Kimi K2.6$5.11
  18. GLM-5.2 (max)$6.06
  19. Qwen3 32B$9.10
  20. Gemini 3.5 Flash$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

20 models; segments: Passive Params, Active Params. Tri-21B-think Preview: 21B, row 19 of 20 in the published order.
  • Passive Params
  • Active Params
  1. DeepSeek V4 Pro (max)1,600B
  2. Kimi K2.61,000B
  3. GLM-5.2 (max)753B
  4. Nemotron 3 Ultra550B
  5. MiniMax-M3428B
  6. gpt-oss-120b (high)117B
  7. Qwen3.6 35B A3B36B
  8. Qwen3 32B32.8B
  9. NVIDIA Nemotron 3 Nano31.6B
  10. Gemma 4 31B30.7B
  11. Qwen3 30B A3B 250730.5B
  12. Qwen3.6 27B27.8B
  13. Qwen3.6 27B27.8B
  14. Gemma 4 26B A4B25.2B
  15. DiffusionGemma 26B A4B25.2B
  16. Devstral Small 224B
  17. Mistral Small 3.124B
  18. Magistral Small 1.224B
  19. Tri-21B-think Preview (this model)21B
  20. gpt-oss-20b (high)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

Tri-21B-think Preview technical specifications
SpecificationValueSource
Model typeOpen weightsfaq
ReasoningYesfaq
Input modalitiesTri-21B-think Preview supports text input.faq
Output modalitiesTri-21B-think Preview supports text output.faq
Context window32k tokensfaq
Open weights / sourceYes, Tri-21B-think Preview is open weights. The model weights are publicly available and can be downloaded for self-hosting.faq
ParametersTri-21B-think Preview has 21 billion parameters.faq
LicenseTri-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

Tri-21B-think Preview was released on February 10, 2026.

Tri-21B-think Preview was created by Trillion Labs.

Tri-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).

Yes, Tri-21B-think Preview is a reasoning model. It uses extended thinking or chain-of-thought reasoning to work through complex problems before providing an answer.

Tri-21B-think Preview supports text input.

Tri-21B-think Preview supports text output.

No, Tri-21B-think Preview does not support image input. It can only process text.

No, Tri-21B-think Preview is not multimodal. It only supports text input.

Tri-21B-think Preview has a context window of 32k tokens. This determines how much text and conversation history the model can process in a single request.

Yes, Tri-21B-think Preview is open weights. The model weights are publicly available and can be downloaded for self-hosting.

Tri-21B-think Preview has 21 billion parameters.

Tri-21B-think Preview is released under the Apache 2.0 license. This license allows commercial use.

Tri-21B-think Preview achieves a score of 14 on the Artificial Analysis Intelligence Index. This composite benchmark evaluates models across reasoning, knowledge, mathematics, and coding.

Tri-21B-think Preview is an open weights model that can be self-hosted.

Tri-21B-think Preview is an open weights model that can be downloaded and self-hosted.

Models