LongCat-2.5

Multimodal
Meituan

LongCat-2.5 is Meituan's native multimodal Mixture-of-Experts model for long-horizon agentic tasks, with roughly 1.6T total parameters and about 48B activated per token. It builds on LongCat-2.0 with stronger native visual understanding and agentic coding, representing images as multi-level discrete tokens (RVQ over native-resolution features) so text and vision share one backbone. Served on the LongCat platform with a 1,048,576-token input window and up to 131,072 output tokens, priced at $0.30/$1.20 per million input/output tokens with $0.006 cached input.

Key Specifications

Parameters
1.6T
Context
1.0M
Release Date
November 1, 2026
Average Score
68.4%

Timeline

Key dates in the model's history
Announcement
November 1, 2026
Last Update
November 3, 2026
Today
November 20, 2026

Technical Specifications

Parameters
1.6T
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$0.30
Output (per 1M tokens)
$1.20
Max Input Tokens
1.0M
Max Output Tokens
131.1K
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
AutomationBench v1.0.6
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; All 600 public tasks, official evaluation code • Self-reported
51.8%
BrowseComp
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Custom lightweight harness with Verify and Discard-All (256K threshold), up to 1,024 exploration rounds; search and browse_with_goal tools only • Self-reported
91.9%
CharXiv-R
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; CharXiv-RQ with tools; max generation 131,972 tokens • Self-reported
90.2%
Claw-Eval
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Pass^3 • Self-reported
78.3%
DeepSWE 1.1
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; mini-swe-agent, max_seq_length 1M; average of three runs • Self-reported
64.6%
FinSearchComp-T3
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95 • Self-reported
63.4%
Frontier Science
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; reported as FrontierScience • Self-reported
77.0%
GDPval-Rubrics
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Reported as GDPVal-AA rubrics • Self-reported
84.8%
Humanity's Last Exam (with tools, text-only)
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Text-only subset with tools, Claude Code v2.1.116, max generation 256K • Self-reported
63.1%
Job Bench
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Main 65; Claude Code harness v2.1.280; official LLM-as-judge protocol with gemini-3.6-flash judge • Self-reported
65.4%
MathVision
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; With tools; max generation 131,972 tokens • Self-reported
96.5%
MCP Atlas
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=0.7 (per footnote 6), top_p=0.95; reasoning effort max, 256K context • Self-reported
83.1%
NL2Repo
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Claude Code harness v2.1.280, max_seq_length 512K • Self-reported
60.5%
PaperBench
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95 • Self-reported
82.1%
Program Bench
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Program Bench 'Almost Resolved' metric as reported by Meituan; Claude Code harness v2.1.280, max_seq_length 512K • Self-reported
25.5%
SkillsBench
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; SkillsBench v1.1 (87 tasks), Codex harness • Self-reported
63.3%
SpreadsheetBench 2
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Claude Code harness v2.1.116, official protocol; Visualization subset judged by gemini-3.6-flash • Self-reported
34.9%
SWE Atlas - Codebase QnA
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Claude Code v2.1.116, Claude Opus 4.5 judge (majority of three); average of three runs • Self-reported
58.3%
SWE Atlas - Test Writing
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Claude Code v2.1.116, Claude Opus 4.5 judge (majority of three); average of three runs • Self-reported
47.8%
Terminal-Bench 2.1
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Claude Code harness v2.1.280, max_seq_length 512K; average of three runs • Self-reported
88.4%
Terminal-Bench 4.0
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Claude Code harness v2.1.280, max_seq_length 512K • Self-reported
38.5%
Video-MME
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Up to 2,048 frames at 1 fps, up to 256 tokens per frame, thinking on • Self-reported
85.2%
Vision2Web
LongCat-2.5 blog evaluation table; enable_thinking=on, temperature=1.0, top_p=0.95; Claude Code v2.1.7, 256K sequence length • Self-reported
78.6%

License & Metadata

License
unknown
Announcement Date
November 1, 2026
Last Updated
November 3, 2026

Compare LongCat-2.5

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.