vs
5:1
overall capability wins (0 ties across 6 scored). Agnes-3.0-Flash Preview leads overall.
- Reasoning4:0
- Specs1:1
Verdict
Agnes-3.0-Flash Preview edges reasoning & knowledge; Agnes-3.0-Flash Preview offers ~2.0x larger context (262K vs 131K).
- Agnes-3.0-Flash Preview edges reasoning & knowledge
- Agnes-3.0-Flash Preview offers ~2.0x larger context (262K vs 131K)
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Agnes-3.0-Flash Preview
- -
- K2 Think V2
- -
Price per 1M tokens
Full comparison
Organization
Agnes AI
MBZUAI / G42 / LLM360
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
131K
Max output
-
-
Parameters
33B
70B
Pricing
Input $/1M
—
—
Output $/1M
—
—
Blended $/1M (3∶1)
-
-
Speed
tok/s
-
-
TTFT (s)
-
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
85.0
71.3
Humanity's Last Exam
—
10.1
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
23.0
18.0
AA-LCR v1.1
68.3
57.0
CritPt
—
—
MMMU-Pro
—
—
IFBench
74.2
62.8
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
—
15.0
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
38.1
—
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 | K2 Think V2 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Agnes AI | MBZUAI / G42 / LLM360 | — |
| 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 | 131K | A +131K |
| Max output | - | - | — |
| Parameters | 33B | 70B | B +37 |
| Pricing | |||
| Input $/1M | — | — | — |
| Output $/1M | — | — | — |
| Blended $/1M (3∶1) | - | - | — |
| Speed | |||
| tok/s | - | - | — |
| TTFT (s) | - | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 85.0 | 71.3 | A +13.8 pts |
| Humanity's Last Exam | — | 10.1 | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 23.0 | 18.0 | A +5.0 pts |
| AA-LCR v1.1 | 68.3 | 57.0 | A +11.3 pts |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | 74.2 | 62.8 | A +11.4 pts |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | — | 15.0 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 38.1 | — | — |
| 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.
K2 Think V2
- Context
- 131K
- Parameters
- 70B
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- Apache 2.0
MBZUAI's Institute of Foundation Models, with G42 and LLM360, fully open 70B dense reasoning model built end-to-end as a 'sovereign' successor to the original K2-Think; no hosted API yet.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.