vs
1:6
overall capability wins (0 ties across 7 scored). MiMo-V2.5-Pro leads overall.
- Reasoning1:3
- Coding0:1
- Specs0:2
Verdict
MiMo-V2.5-Pro leads coding (widest gap: +12.5 on SciCode); MiMo-V2.5-Pro offers ~3.8x larger context (1M vs 262K).
- MiMo-V2.5-Pro leads coding (widest gap: +12.5 on SciCode)
- MiMo-V2.5-Pro 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-Pro
- -
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 → text
Specs
Context window
262K
1M
Max output
-
-
Parameters
33B
1.0T (42B act.)
Pricing
Input $/1M
—
—
Output $/1M
—
—
Blended $/1M (3∶1)
-
-
Speed
tok/s
-
-
TTFT (s)
-
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
85.0
86.6
Humanity's Last Exam
—
35.7
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
23.0
22.4
AA-LCR v1.1
68.3
79.7
CritPt
—
4.0
MMMU-Pro
—
—
IFBench
74.2
79.9
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
—
65.2
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
38.1
50.6
CursorBench
—
—
SWE-Rebench
—
44.4
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
1,475
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
1,467
| Metric | Agnes-3.0-Flash Preview | MiMo-V2.5-Pro | 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 → text | — |
| Specs | |||
| Context window | 262K | 1M | B +738K |
| Max output | - | - | — |
| Parameters | 33B | 1.0T (42B act.) | B +987 |
| Pricing | |||
| Input $/1M | — | — | — |
| Output $/1M | — | — | — |
| Blended $/1M (3∶1) | - | - | — |
| Speed | |||
| tok/s | - | - | — |
| TTFT (s) | - | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 85.0 | 86.6 | B +1.5 pts |
| Humanity's Last Exam | — | 35.7 | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 23.0 | 22.4 | A +0.6 pts |
| AA-LCR v1.1 | 68.3 | 79.7 | B +11.4 pts |
| CritPt | — | 4.0 | — |
| MMMU-Pro | — | — | — |
| IFBench | 74.2 | 79.9 | B +5.7 pts |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | — | 65.2 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 38.1 | 50.6 | B +12.5 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | 44.4 | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | 1,475 | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | 1,467 | — |
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-Pro
- Context
- 1M
- Parameters
- 1.0T (42B act.)
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- MIT
Xiaomi's flagship open-weight MoE model (1.02T total / 42B active parameters) in the MiMo-V2.5 series, tuned for coding, reasoning, and long-horizon agent tasks with a 1M-token context window, released under the MIT license.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.