vs
5:2
overall capability wins (0 ties across 7 scored). K2 Horizon 375B A23B leads overall.
- Reasoning4:1
- Coding1:0
- Specs0:1
Verdict
K2 Horizon 375B A23B leads coding (widest gap: +0.3 on SciCode); MiniMax M2.5 offers ~1.9x larger context (1M vs 524K).
- K2 Horizon 375B A23B leads coding (widest gap: +0.3 on SciCode)
- MiniMax M2.5 offers ~1.9x larger context (1M vs 524K)
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- K2 Horizon 375B A23B
- -
- MiniMax M2.5
- $0.60
Price per 1M tokens
Full comparison
Organization
MBZUAI Institute of Foundation Models
MiniMax
Family
Other
Other
License
Apache 2.0
Modified MIT
Open weights
Yes
Yes
Release
Sep 3, 2026
Feb 15, 2026
Knowledge cutoff
-
-
API / provider
No primary API price listed
MiniMax
Modalities
text → text
text → text
Specs
Context window
524K
1M
Max output
-
64K
Parameters
375B (23B act.)
—
Pricing
Input $/1M
—
$0.30
Output $/1M
—
$1.20
Blended $/1M (3∶1)
-
$0.52
Speed
tok/s
-
85
TTFT (s)
-
0.35
Reasoning
MMLU-Pro
—
—
GPQA Diamond
87.3
84.8
Humanity's Last Exam
32.0
20.5
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
18.2
26.2
AA-LCR v1.1
80.0
73.3
CritPt
4.6
1.1
MMMU-Pro
—
—
IFBench
—
71.6
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
75.8
SWE-bench Pro
—
—
SWE-bench Multilingual
—
68.3
LiveCodeBench
—
—
Terminal-Bench 2.1
71.9
—
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
42.9
42.6
CursorBench
—
—
SWE-Rebench
—
47.7
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
1,384
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
1,390
| Metric | K2 Horizon 375B A23B | MiniMax M2.5 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | MBZUAI Institute of Foundation Models | MiniMax | — |
| Family | Other | Other | — |
| License | Apache 2.0 | Modified MIT | — |
| Open weights | Yes | Yes | — |
| Release | Sep 3, 2026 | Feb 15, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | No primary API price listed | MiniMax | — |
| Modalities | text → text | text → text | — |
| Specs | |||
| Context window | 524K | 1M | B +476K |
| Max output | - | 64K | — |
| Parameters | 375B (23B act.) | — | — |
| Pricing | |||
| Input $/1M | — | $0.30 | — |
| Output $/1M | — | $1.20 | — |
| Blended $/1M (3∶1) | - | $0.52 | — |
| Speed | |||
| tok/s | - | 85 | — |
| TTFT (s) | - | 0.35 | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 87.3 | 84.8 | A +2.5 pts |
| Humanity's Last Exam | 32.0 | 20.5 | A +11.5 pts |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 18.2 | 26.2 | B +8.0 pts |
| AA-LCR v1.1 | 80.0 | 73.3 | A +6.7 pts |
| CritPt | 4.6 | 1.1 | A +3.5 pts |
| MMMU-Pro | — | — | — |
| IFBench | — | 71.6 | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | 75.8 | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | 68.3 | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | 71.9 | — | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 42.9 | 42.6 | A +0.3 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | 47.7 | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | 1,384 | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | 1,390 | — |
K2 Horizon 375B A23B
- Context
- 524K
- Parameters
- 375B (23B act.)
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- Apache 2.0
Flagship model (375B MoE / 23B active) in the Institute of Foundation Models' K2 Horizon fleet, released fully open — weights, code, training data, and methodology — under Apache 2.0.
MiniMax M2.5
- Context
- 1M / 64K out
- Parameters
- —
- Price
- $0.30 / $1.20
- Speed
- 85 tok/s · 0.35s TTFT
- Modalities
- text → text
- License
- Modified MIT
Open-ish MiniMax MoE with long context and competitive Chinese-market pricing.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.