vs
3:4
overall capability wins (0 ties across 7 scored). Agnes-3.0-Flash Preview leads overall.
- Reasoning3:1
- Coding0:1
- Specs0:2
Verdict
MiMo-V2.5 leads coding (widest gap: +5.8 on SciCode); Agnes-3.0-Flash Preview edges reasoning & knowledge; MiMo-V2.5 offers ~3.8x larger context (1M vs 262K).
- MiMo-V2.5 leads coding (widest gap: +5.8 on SciCode)
- Agnes-3.0-Flash Preview edges reasoning & knowledge
- MiMo-V2.5 offers ~3.8x larger context (1M vs 262K)
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Agnes-3.0-Flash Preview
- -
- MiMo-V2.5
- -
Price per 1M tokens
Full comparison
Organization
Agnes AI
Xiaomi
Family
Other
Other
License
Apache 2.0
MIT
Open weights
Yes
Yes
Release
Sep 13, 2026
Apr 27, 2026
Knowledge cutoff
-
-
API / provider
No primary API price listed
No primary API price listed
Modalities
text, image, video → text
text, image, audio, video → text
Specs
Context window
262K
1M
Max output
-
-
Parameters
33B
310B (15B act.)
Pricing
Input $/1M
—
—
Output $/1M
—
—
Blended $/1M (3∶1)
-
-
Speed
tok/s
-
-
TTFT (s)
-
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
85.0
84.9
Humanity's Last Exam
—
27.2
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
23.0
16.8
AA-LCR v1.1
68.3
73.0
CritPt
—
3.7
MMMU-Pro
—
75.4
IFBench
74.2
67.1
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
—
63.7
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
38.1
43.9
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
1,437
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
1,434
| Metric | Agnes-3.0-Flash Preview | MiMo-V2.5 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Agnes AI | Xiaomi | — |
| Family | Other | Other | — |
| License | Apache 2.0 | MIT | — |
| Open weights | Yes | Yes | — |
| Release | Sep 13, 2026 | Apr 27, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | No primary API price listed | No primary API price listed | — |
| Modalities | text, image, video → text | text, image, audio, video → text | — |
| Specs | |||
| Context window | 262K | 1M | B +738K |
| Max output | - | - | — |
| Parameters | 33B | 310B (15B act.) | B +277 |
| Pricing | |||
| Input $/1M | — | — | — |
| Output $/1M | — | — | — |
| Blended $/1M (3∶1) | - | - | — |
| Speed | |||
| tok/s | - | - | — |
| TTFT (s) | - | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 85.0 | 84.9 | A +0.1 pts |
| Humanity's Last Exam | — | 27.2 | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 23.0 | 16.8 | A +6.2 pts |
| AA-LCR v1.1 | 68.3 | 73.0 | B +4.7 pts |
| CritPt | — | 3.7 | — |
| MMMU-Pro | — | 75.4 | — |
| IFBench | 74.2 | 67.1 | A +7.1 pts |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | — | 63.7 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 38.1 | 43.9 | B +5.8 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | 1,437 | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | 1,434 | — |
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.
MiMo-V2.5
- Context
- 1M
- Parameters
- 310B (15B act.)
- Price
- —
- Speed
- —
- Modalities
- text, image, audio, video → text
- License
- MIT
Xiaomi's omnimodal sibling to MiMo-V2.5-Pro (310B total / 15B active parameters), understanding text, image, video, and audio within a 1M-token context window, released open-weight under the MIT license.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.