vs
4:2
overall capability wins (1 ties across 7 scored). Agnes-3.0-Flash Preview leads overall.
- Reasoning4:0
- Coding0:1
- Specs0:1
Verdict
Trinity Large Thinking leads coding (widest gap: +2.5 on SciCode); Agnes-3.0-Flash Preview edges reasoning & knowledge.
- Trinity Large Thinking leads coding (widest gap: +2.5 on SciCode)
- Agnes-3.0-Flash Preview edges reasoning & knowledge
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Agnes-3.0-Flash Preview
- -
- Trinity Large Thinking
- -
Price per 1M tokens
Full comparison
Organization
Agnes AI
Arcee AI
Family
Other
Other
License
Apache 2.0
Apache 2.0
Open weights
Yes
Yes
Release
Sep 13, 2026
Jan 27, 2026
Knowledge cutoff
-
-
API / provider
No primary API price listed
No primary API price listed
Modalities
text, image, video → text
text → text
Specs
Context window
262K
262K
Max output
-
-
Parameters
33B
400B (13B act.)
Pricing
Input $/1M
—
—
Output $/1M
—
—
Blended $/1M (3∶1)
-
-
Speed
tok/s
-
-
TTFT (s)
-
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
85.0
75.2
Humanity's Last Exam
—
15.8
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
23.0
22.5
AA-LCR v1.1
68.3
38.0
CritPt
—
0.9
MMMU-Pro
—
—
IFBench
74.2
56.3
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
—
20.6
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
38.1
40.6
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
1,237
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
1,369
| Metric | Agnes-3.0-Flash Preview | Trinity Large Thinking | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Agnes AI | Arcee AI | — |
| Family | Other | Other | — |
| License | Apache 2.0 | Apache 2.0 | — |
| Open weights | Yes | Yes | — |
| Release | Sep 13, 2026 | Jan 27, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | No primary API price listed | No primary API price listed | — |
| Modalities | text, image, video → text | text → text | — |
| Specs | |||
| Context window | 262K | 262K | tie |
| Max output | - | - | — |
| Parameters | 33B | 400B (13B act.) | B +367 |
| Pricing | |||
| Input $/1M | — | — | — |
| Output $/1M | — | — | — |
| Blended $/1M (3∶1) | - | - | — |
| Speed | |||
| tok/s | - | - | — |
| TTFT (s) | - | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 85.0 | 75.2 | A +9.8 pts |
| Humanity's Last Exam | — | 15.8 | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 23.0 | 22.5 | A +0.5 pts |
| AA-LCR v1.1 | 68.3 | 38.0 | A +30.3 pts |
| CritPt | — | 0.9 | — |
| MMMU-Pro | — | — | — |
| IFBench | 74.2 | 56.3 | A +17.9 pts |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | — | 20.6 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 38.1 | 40.6 | B +2.5 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | 1,237 | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | 1,369 | — |
Agnes-3.0-Flash Preview
- Context
- 262K
- Parameters
- 33B
- Price
- —
- Speed
- —
- Modalities
- text, image, video → text
- License
- Apache 2.0
Agnes AI's (Sapiens AI, Singapore) first open-weight release: a 33B-parameter hybrid-attention model (54 gated delta-rule recurrent layers alternating with 18 global-attention layers, 3:1) under Apache 2.0, with a 262K-token context window and text/image/video input. This open-weight Preview checkpoint is distinct from Agnes AI's production "Agnes 3.0 Flash" API model, which runs a different checkpoint and is not open-weight; the two should not be conflated.
Trinity Large Thinking
- Context
- 262K
- Parameters
- 400B (13B act.)
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- Apache 2.0
Arcee AI's reasoning-tuned Trinity Large (400B total / 13B active MoE, 4-of-256 experts), Apache 2.0 open weights, larger sibling to Trinity Mini; released Jan 27, 2026.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.