vs
5:0
overall capability wins (1 ties across 6 scored). Agnes-3.0-Flash Preview leads overall.
- Reasoning3:0
- Coding1:0
- Specs1:0
Verdict
Agnes-3.0-Flash Preview leads coding (widest gap: +13.9 on SciCode).
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Agnes-3.0-Flash Preview
- -
- Ling-3.0 Tiny
- -
Price per 1M tokens
Full comparison
Organization
Agnes AI
InclusionAI
Family
Other
Other
License
Apache 2.0
MIT
Open weights
Yes
Yes
Release
Sep 13, 2026
Aug 17, 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
-
33K
Parameters
33B
7.9B (1.3B act.)
Pricing
Input $/1M
—
—
Output $/1M
—
—
Blended $/1M (3∶1)
-
-
Speed
tok/s
-
-
TTFT (s)
-
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
85.0
73.4
Humanity's Last Exam
—
9.3
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
23.0
8.5
AA-LCR v1.1
68.3
60.3
CritPt
—
—
MMMU-Pro
—
—
IFBench
74.2
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
—
27.7
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
38.1
24.2
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
—
| Metric | Agnes-3.0-Flash Preview | Ling-3.0 Tiny | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Agnes AI | InclusionAI | — |
| Family | Other | Other | — |
| License | Apache 2.0 | MIT | — |
| Open weights | Yes | Yes | — |
| Release | Sep 13, 2026 | Aug 17, 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 | - | 33K | — |
| Parameters | 33B | 7.9B (1.3B act.) | A +25 |
| Pricing | |||
| Input $/1M | — | — | — |
| Output $/1M | — | — | — |
| Blended $/1M (3∶1) | - | - | — |
| Speed | |||
| tok/s | - | - | — |
| TTFT (s) | - | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 85.0 | 73.4 | A +11.6 pts |
| Humanity's Last Exam | — | 9.3 | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 23.0 | 8.5 | A +14.5 pts |
| AA-LCR v1.1 | 68.3 | 60.3 | A +8.0 pts |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | 74.2 | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | — | 27.7 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 38.1 | 24.2 | A +13.9 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | — | — |
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.
Ling-3.0 Tiny
- Context
- 262K / 33K out
- Parameters
- 7.9B (1.3B act.)
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- MIT
Compact on-device member of InclusionAI's Ling 3.0 family (7.9B total / 1.3B active MoE, MIT) with a hybrid KDA/MLA attention stack, native 256K context, and switchable thinking / instant modes.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.