vs
6:4
overall capability wins (0 ties across 10 scored). K2 Horizon 375B A23B leads overall.
- Reasoning3:2
- Coding2:0
- Tool use1:0
- Composite indices2:0
- Specs0:2
Verdict
K2 Horizon 375B A23B leads coding (widest gap: +21.7 on TermBench); LongCat-2.0 offers ~1.9x larger context (1M vs 524K).
- K2 Horizon 375B A23B leads coding (widest gap: +21.7 on TermBench)
- LongCat-2.0 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
- -
- LongCat-2.0
- -
Price per 1M tokens
Full comparison
Organization
MBZUAI Institute of Foundation Models
Meituan
Family
Other
Other
License
Apache 2.0
MIT
Open weights
Yes
Yes
Release
Sep 3, 2026
Jun 29, 2026
Knowledge cutoff
-
-
API / provider
No primary API price listed
No primary API price listed
Modalities
text → text
text → text
Specs
Context window
524K
1M
Max output
-
64K
Parameters
375B (23B act.)
1.6T (48B act.)
Pricing
Input $/1M
—
—
Output $/1M
—
—
Blended $/1M (3∶1)
-
-
Speed
tok/s
-
55
TTFT (s)
-
0.5
Reasoning
MMLU-Pro
—
—
GPQA Diamond
87.3
78.0
Humanity's Last Exam
32.0
33.7
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
18.2
29.6
AA-LCR v1.1
80.0
65.0
CritPt
4.6
2.6
MMMU-Pro
—
—
IFBench
—
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
71.9
50.2
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
42.9
36.3
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
—
| Metric | K2 Horizon 375B A23B | LongCat-2.0 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | MBZUAI Institute of Foundation Models | Meituan | — |
| Family | Other | Other | — |
| License | Apache 2.0 | MIT | — |
| Open weights | Yes | Yes | — |
| Release | Sep 3, 2026 | Jun 29, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | No primary API price listed | No primary API price listed | — |
| Modalities | text → text | text → text | — |
| Specs | |||
| Context window | 524K | 1M | B +476K |
| Max output | - | 64K | — |
| Parameters | 375B (23B act.) | 1.6T (48B act.) | B +1K |
| Pricing | |||
| Input $/1M | — | — | — |
| Output $/1M | — | — | — |
| Blended $/1M (3∶1) | - | - | — |
| Speed | |||
| tok/s | - | 55 | — |
| TTFT (s) | - | 0.5 | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 87.3 | 78.0 | A +9.3 pts |
| Humanity's Last Exam | 32.0 | 33.7 | B +1.7 pts |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 18.2 | 29.6 | B +11.4 pts |
| AA-LCR v1.1 | 80.0 | 65.0 | A +15.0 pts |
| CritPt | 4.6 | 2.6 | A +2.0 pts |
| MMMU-Pro | — | — | — |
| IFBench | — | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | 71.9 | 50.2 | A +21.7 pts |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 42.9 | 36.3 | A +6.6 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | — | — |
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.
LongCat-2.0
- Context
- 1M / 64K out
- Parameters
- 1.6T (48B act.)
- Price
- —
- Speed
- 55 tok/s · 0.5s TTFT
- Modalities
- text → text
- License
- MIT
1.6T MoE trained on domestic Chinese chips - open MIT weights focused on agentic coding.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.