Agentic coding focus
Shape coding tasks, repo analysis, terminal reasoning, and multi-step software work around LongCat 2.0 capabilities.
LongCat 2.0 Online
LongCat 2.0 for online agentic coding workflows
LongCat 2.0 Online is a clean, practical workspace for exploring LongCat 2.0, the open-source MoE model built for agentic coding and long-context engineering workflows.
longcat20.online/session
model: LongCat-2.0 mode: agentic-coding context: 1M tokens active_params: dynamic tasks: - inspect repository - reason across files - plan terminal actions - produce implementation brief
1.6T
total MoE parameters
1M
native context window
33B-56B
dynamic active parameters
MIT
open-source license
Live Playground
Try LongCat 2.0 directly from the page
Send a prompt through the server-side LongCat API proxy and review the model response without exposing the API key in the browser.
Input
LongCat-2.0Output
The generated answer will appear here after you run the model.
Product Direction
A focused interface for model research and coding operations
Long-context ready
Designed for workflows that need broad project context, technical documents, logs, and implementation notes in one place.
Evaluation-aware
Track model behavior against coding and terminal-style benchmarks without hiding the practical tradeoffs.
Online research hub
Bring release notes, model facts, and workflow guidance into a focused LongCat 2.0 Online SaaS experience.
Built around the questions evaluators actually ask
Compare LongCat 2.0 behavior across coding prompts
Draft agentic workflows for repo-level implementation tasks
Collect model notes, links, and benchmark context
Prepare technical briefs for teams evaluating open models
Official Benchmark Results
LongCat 2.0 official evaluation charts
These images are local screenshots of the benchmark charts on the official LongCat 2.0 blog page. The official page groups LongCat 2.0 results into code-agent, general-agent, and search-agent evaluations, and compares them with leading proprietary models.
The official page states that values marked with an asterisk are external reported metrics, while the remaining values are measured in-house under a unified harness. LongCat 2.0 Online keeps that source boundary visible instead of presenting the charts as our own benchmark run.
Verify on the official LongCat 2.0 blog
Official table capture
This table screenshot comes from the same official blog section and preserves the benchmark grouping and footnote context.
FAQ
What is LongCat 2.0 Online?
LongCat 2.0 Online is a SaaS site concept for exploring LongCat 2.0 information, workflows, and model evaluation context in a focused web interface.
Is this the official LongCat 2.0 site?
No. This site is an independent online workspace and guide. It can link to official project pages and research references for source material.
Why emphasize agentic coding?
LongCat 2.0 is positioned around coding-agent and terminal-oriented tasks, so the product language focuses on practical software engineering workflows.
Source-based LongCat 2.0 guide
LongCat 2.0 facts for model evaluators
LongCat 2.0 Online summarizes public LongCat 2.0 facts for technical readers who want a compact evaluation brief before opening the original sources. The section below separates architecture, context, benchmark signals, source boundaries, and evaluation cautions so readers can scan LongCat 2.0 claims without working through a long article.
Model scale
1.6T MoE
LongCat 2.0 is described in official release material as a 1.6 trillion parameter mixture-of-experts model. LongCat 2.0 uses dynamic activation rather than one fixed active-parameter count.
Active parameters
33B-56B
The official material reports about 33 billion to 56 billion active parameters, with approximately 48 billion active parameters on average. LongCat 2.0 should be compared as a sparse model, not as a dense 1.6T model.
Context window
1M tokens
LongCat 2.0 is reported with a native one million token context window. The release connects that context length with Large-scale Sparse Attention, described as LSA.
License
MIT
LongCat 2.0 is presented as an open-source model under the MIT License. LongCat 2.0 Online still points readers to official sources for exact release and license details.
Reported benchmark groups
Coding and terminal signal
- LongCat 2.0 is reported at 59.5 on SWE-bench Pro.
- LongCat 2.0 is reported at 70.8 on Terminal-Bench 2.1.
- These LongCat 2.0 scores are useful starting points, not a guarantee for every repository workflow.
Search and browsing signal
- LongCat 2.0 is reported at 78.8 on RWSearch.
- LongCat 2.0 is reported at 73.2 on FORTE.
- LongCat 2.0 is reported at 79.9 on BrowseComp.
Evaluator checklist
- Verify LongCat 2.0 model size against the official release.
- Verify LongCat 2.0 context length and LSA wording before reusing claims.
- Verify LongCat 2.0 benchmark values from the linked official source.
- Treat LongCat 2.0 Online as an independent guide, not the official LongCat 2.0 project.
- Test LongCat 2.0 inside local prompts, tools, latency limits, and cost assumptions.
- Keep LongCat 2.0 source links close to claims that readers may want to audit.
How to read this
LongCat 2.0 Online keeps LongCat 2.0 architecture notes, LongCat 2.0 benchmark notes, LongCat 2.0 source links, and LongCat 2.0 evaluation cautions together. The goal is a structured LongCat 2.0 reference, not a replacement for official LongCat 2.0 material.