Language modelActive

Trinity Large Thinking

Arcee AI

Released: April 1, 2026
Data date: September 8, 2026

Speed: 319.5 tokens/s
Output throughput
Price: $0.25 · $0.8
Input · Output / 1M tokens
Input: Text
Output: Text

Trinity Large Thinking exposes its reasoning trace as visible reasoning blocks. The open sparse MoE has 398 billion total parameters, activates 13 billion, routes four of 256 experts per token alongside shared experts, and starts with six dense layers.

Arcee’s model card lists 524,288 tokens after an extension, while the current executable configuration sets max_position_embeddings to 262,144 and shows no matching RoPE scaling. OpenMDW-1.1 permits use and modification of the weights.

Specifications and access

Specification	Value and source
Weights	Open weightsSource
Access	API, LocalSource
Parameters	398BSource
Active parameters	13BSource
Context window	262,144 tokensSource
License	OpenMDW-1.1Source
Parameters	398 BSource
Active parameters	13 BSource
Architecture	sparse MoE AfmoeForCausalLM, 256 experts plus one shared, four active experts, six dense layersSource
Pretraining context	8,192 tokensSource
Advertised context window	524,288 tokens (Vendor-advertised context length after extension. The exact current config.json separately declares max_position_embeddings=262144 with rope_scaling=null.)Source
Training tokens	17TSource
Input	TextSource
Output	TextSource
Reasoning mode	reasoning-optimized; explicit <think>...</think> tracesSource
Tool calling	YesSource
Model configuration context	262,144 tokens (Executable config value; distinct from the card's advertised 512K context after extension.)Source

Published benchmarks

CorpFin v2

Overall

1.Claude Opus 573.19%

2.Claude Fable 571.83%

3.Kimi K371.56%

4.Inkling Small69.62%

5.Inkling68.57%

6.Grok 4.368.53%

7.GPT-5.568.42%

8.Kimi K2.568.26%

9.MiniMax M368.1%

10.Grok 4.567.41%

11.Claude Sonnet 566.98%

12.Kimi K2.666.74%

13.Claude Opus 4.866.71%

14.Gemini 3 Flash Preview66.43%

15.Grok 4.666.16%

16.GLM-5.266.12%

17.Claude Opus 4.766.08%

18.Qwen 3.8 Max 090265.85%

19.Nemotron 3 Ultra65.46%

20.Claude Sonnet 4.665.31%

20.GPT-5.6 Terra65.31%

22.GPT-5.465.27%

23.Gemini 3.5 Flash64.69%

24.Gemini 3.1 Pro Preview64.49%

25.GLM-5.164.45%

26.GPT-5.6 Sol64.38%

27.GPT-5.6 Luna64.22%

28.Qwen 3.7 Max63.71%

29.Grok 4.20 Reasoning63.68%

30.Gemini 3.6 Flash63.33%

31.GPT-4.163.05%

32.GLM-562.9%

33.Qwen 3.6 27B62.32%

34.Ling 3.0 Flash61.93%

35.MiMo-V2.5-Pro61.42%

36.DeepSeek-V4-Pro61.38%

37.Claude Opus 4.561.31%

38.GPT-5.4 nano61.19%

38.MiniMax M2.761.19%

40.GPT-561.07%

41.Mistral Large 361.03%

42.GLM-4.560.96%

43.GPT-5.4 mini60.92%

44.Claude Sonnet 4.560.8%

44.Gemini 2.5 Pro60.8%

46.Gemini 3.5 Flash-Lite60.68%

47.Kimi K2 Thinking60.57%

48.Claude Haiku 4.560.3%

49.GPT-5 mini60.18%

50.MiMo-V2.559.91%

51.o359.71%

52.MiniMax M2.559.6%

53.Gemini 3.1 Flash-Lite59.36%

54.o4-mini58.97%

55.MiniMax M2.158.9%

56.Mistral Medium 3.558.78%

57.GPT OSS 120B58.24%

58.GPT-4.1 mini57.93%

59.GLM-4.656.84%

60.DeepSeek-V3 032454.74%

61.Trinity Large Thinking54.66%

62.Nemotron 3.5 Lightning54.24%

63.DeepSeek-R154.12%

64.GPT OSS 20B53.15%

65.DeepSeek-V352.49%

66.DeepSeek-V3.151.48%

67.Kimi K2 071150.39%

68.DeepSeek-V3.247.94%

69.Llama 4 Scout46.78%

70.GLM-4.746.39%

71.Command A45.96%

72.GPT-4o mini45.46%

73.o3-mini45.3%

74.Mistral Small 3.144.17%

75.GPT-4.1 nano42.08%

higher is betterAs of September 29, 2026Scale: 0% - 100%

Source: Vals AI

gradually.ai

Methodology and context

Comparison cohort: vals-corp-fin-v2:2:overall
Metric: accuracy
Scale: 0-100
Participants: 75
Methodology and context: Source model ID: arcee-ai/trinity-large-thinking. Source model name: Trinity Large Thinking. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-08-12. Latency in seconds: 12.482. Cost per task in USD: 0.01939. Source ID: vals-ai. Evidence capture ID: vals-corp_fin_v2-2026-09-29-83186cfafe52. Source payload hash: 83186cfafe526f0b8fe2ae86af05e8209fe799c8bfa48292ef1fef5db66650ee. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

Terminal-Bench 2.0

Overall

1.GPT-5.573.2%

2.Claude Opus 4.870.04%

3.Claude Opus 4.768.54%

4.Gemini 3.1 Pro Preview67.42%

4.Gemini 3.5 Flash67.42%

6.GPT-5.3-Codex64.05%

7.Claude Sonnet 4.659.55%

8.Qwen 3.7 Max59.18%

9.Claude Opus 4.558.43%

9.GPT-5.458.43%

11.Kimi K2.657.3%

12.DeepSeek-V4-Pro56.18%

13.GLM-5.153.93%

14.Gemini 3 Flash Preview51.69%

15.GLM-549.44%

16.MiniMax M2.747.19%

17.MiniMax M346.07%

18.GPT-5.4 mini44.94%

18.Qwen 3.6 27B44.94%

20.Grok 4.343.45%

21.MiniMax M2.541.57%

22.Grok 4.20 Reasoning40.45%

22.Kimi K2.540.45%

24.GPT-5.4 nano39.89%

25.Gemma 4 31B39.33%

26.GLM-4.738.2%

27.GPT-537.08%

27.Kimi K2 Thinking37.08%

27.MiniMax M2.137.08%

30.DeepSeek-V3.234.83%

31.Gemini 2.5 Pro30.34%

31.Mistral Medium 3.530.34%

33.GLM-4.628.09%

34.GPT-5 mini26.97%

35.Kimi K2 071125.84%

36.Gemini 3.1 Flash-Lite24.72%

37.DeepSeek-V3.122.47%

38.GPT OSS 120B19.1%

39.Trinity Large Thinking17.98%

40.Mistral Small 416.85%

41.GPT-4.114.61%

42.Mistral Large 38.99%

43.Command A2.25%

higher is betterAs of September 29, 2026Scale: 0% - 100%

Source: Vals AI

gradually.ai

Methodology and context

Comparison cohort: vals-terminal-bench-2:2.0:overall
Metric: accuracy
Scale: 0-100
Participants: 43
Methodology and context: Source model ID: arcee-ai/trinity-large-thinking. Source model name: Trinity Large Thinking. Evaluator: Vals AI. Dataset type: public. Leaderboard date: 2026-06-04. Latency in seconds: 1015.224. Cost per task in USD: 0.325546. Source ID: vals-ai. Evidence capture ID: vals-terminal-bench-2-2026-09-29-020c4135b825. Source payload hash: 020c4135b82502d8cc0324b9b7d9474024a86da8f4770c78743bd13103cdfa0c. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

CaseLaw v2: Overall (Show measurement, test conditions, and source)

Source value: 57.878
Score: 57.878
Metric: accuracy
Unit: %
Benchmark version: 2
Category: legal
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-case-law-v2:2:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: arcee-ai/trinity-large-thinking. Source model name: Trinity Large Thinking. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-05-04. Latency in seconds: 14.554. Cost per task in USD: 0.008867. Source ID: vals-ai. Evidence capture ID: vals-case_law_v2-2026-09-29-cd282b756eb4. Source payload hash: cd282b756eb41a617c01694c5ed3b3d1872a9f417d8aac675370f4943c6c7c63. Parser version: vals-astro-v2. Imported at: 2026-09-29
Context: The source dataset marks this benchmark as archived. The result remains visible as historical evidence and is not interchangeable with an active leaderboard.

Source Vals AI

IFBench (Show measurement, test conditions, and source)

Source value: 52.3
Score: 52.3
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.

Source Arcee Trinity-Large-Thinking model card

GPQA-Diamond (Show measurement, test conditions, and source)

Source value: 76.3
Score: 76.3
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.

Source Arcee Trinity-Large-Thinking model card

Tau2-Airline (Show measurement, test conditions, and source)

Source value: 88
Score: 88
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.

Source Arcee Trinity-Large-Thinking model card

Tau2-Telecom (Show measurement, test conditions, and source)

Source value: 94.7
Score: 94.7
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.

Source Arcee Trinity-Large-Thinking model card

PinchBench (Show measurement, test conditions, and source)

Source value: 91.9
Score: 91.9
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.

Source Arcee Trinity-Large-Thinking model card

AIME25 (Show measurement, test conditions, and source)

Source value: 96.3
Score: 96.3
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.

Source Arcee Trinity-Large-Thinking model card

BCFLv4 (Show measurement, test conditions, and source)

Source value: 70.1
Score: 70.1
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.
Context: Benchmark name is spelled BCFLv4 in the exact card; retained verbatim.

Source Arcee Trinity-Large-Thinking model card

MMLU-Pro (Show measurement, test conditions, and source)

Source value: 83.4
Score: 83.4
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.

Source Arcee Trinity-Large-Thinking model card

SWE-bench Verified (Show measurement, test conditions, and source)

Source value: 63.2
Score: 63.2
Metric: percent
Unit: %
Category: mixed
Direction: Higher is better
Source type: vendor-reported
Evaluator: Arcee Trinity-Large-Thinking model card
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Vendor comparison table; all models evaluated in mini-swe-agent-v2 where stated; exact row settings otherwise follow the card.

Source Arcee Trinity-Large-Thinking model card

Recorded measurement details

CorpFin v2: Overall (Show measurement, test conditions, and source)

Source value: 54.662
Score: 54.662
Metric: accuracy
Unit: %
Benchmark version: 2
Category: finance
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-corp-fin-v2:2:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: arcee-ai/trinity-large-thinking. Source model name: Trinity Large Thinking. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-08-12. Latency in seconds: 12.482. Cost per task in USD: 0.01939. Source ID: vals-ai. Evidence capture ID: vals-corp_fin_v2-2026-09-29-83186cfafe52. Source payload hash: 83186cfafe526f0b8fe2ae86af05e8209fe799c8bfa48292ef1fef5db66650ee. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

Terminal-Bench 2.0: Overall (Show measurement, test conditions, and source)

Source value: 17.978
Score: 17.978
Metric: accuracy
Unit: %
Benchmark version: 2.0
Category: agentic
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-terminal-bench-2:2.0:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: arcee-ai/trinity-large-thinking. Source model name: Trinity Large Thinking. Evaluator: Vals AI. Dataset type: public. Leaderboard date: 2026-06-04. Latency in seconds: 1015.224. Cost per task in USD: 0.325546. Source ID: vals-ai. Evidence capture ID: vals-terminal-bench-2-2026-09-29-020c4135b825. Source payload hash: 020c4135b82502d8cc0324b9b7d9474024a86da8f4770c78743bd13103cdfa0c. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

Page 1 of 2

Sources and data date

Type	Evidence and data date
Speed	Median across serving providers. Output speed after the first streaming chunk, not total response time. Retrieved September 8, 2026. Source: Artificial Analysis · Artificial Analysis (retrieved September 8, 2026)Speed measurement methodology
Price	Base API prices without caching or batch discounts. Higher context tiers and other rates are listed under Costs. Source: Pricing · API pricing
Input	Documented modalities. Retrieved September 8, 2026. Source: Model documentation · Model metadata (retrieved September 8, 2026) · Arcee AI Trinity Large Thinking model card · Editorial description reviewed October 2, 2026
Output	Documented modalities. Retrieved September 8, 2026. Source: Model documentation · Model metadata (retrieved September 8, 2026) · Arcee AI Trinity Large Thinking model card · Editorial description reviewed October 2, 2026
Research status	Research date September 8, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources	knowledge-cutoff Original wording The exact Trinity-Large-Thinking card does not publish a knowledge cutoff.
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Model metadata (retrieved September 8, 2026) · Trinity Large Thinking executable configuration · Editorial description reviewed October 2, 2026
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Creative Commons Attribution 4.0

Trinity Large Thinking

Specifications and access

Published benchmarks

CorpFin v2

Terminal-Bench 2.0

Measurements without matching peer values

Recorded measurement details

Our tests

SVG short test

Trinity Large Thinking

Pelican on a bicycle

Popularity

Published prices

Sources and data date