Changelog

This site was initialized as a content-first site for LongCat 2.0 research and evaluation. Information on the first release was organized around model size, long context, dynamic active parameters, open-source status, and published benchmark references. This site started with a simple light interface.

This site replaced template branding with LongCat 2.0 branding. It also replaced the old favicon, logo mark, OG image, and app metadata. It uses LC as the mark because LongCat shortens naturally to LC.

This site removed template blog and docs routes from public navigation. This site did this to avoid publishing unrelated starter-template material. This site can add a research library later, but only when the pages contain LongCat 2.0 material.

LongCat 2.0 Online uses English as the default locale and avoids an /en route for the default page. LongCat 2.0 Online also runs in light mode without a visible mode selector. LongCat 2.0 Online made those choices because the first release is meant to be simple and stable.

This site expanded content after the first initialization. It now targets longer content pages with a deliberate LongCat 2.0 keyword density. It still treats source accuracy as more important than adding unsupported statements.

This site passed a production build after the cleanup work. It still warns in local builds when OAuth keys are not configured. This site keeps those warnings separate from content quality because authentication secrets are deployment configuration.

LongCat 2.0 changelog entries should make LongCat 2.0 changes easy to audit. LongCat 2.0 content updates are listed separately from design updates.

LongCat 2.0 Online uses LongCat 2.0 in update notes where the change affects LongCat 2.0 positioning, source handling, or keyword coverage.

LongCat 2.0 is described by its official release material as a 1.6 trillion parameter mixture-of-experts model. The model uses dynamic activation rather than a fixed active-parameter budget. It is reported with about 33 billion to 56 billion active parameters, with approximately 48 billion active parameters on average across contexts.

The release is presented as an open-source model under the MIT License. It is positioned for agentic coding, agentic search, and long-context reasoning. The model is therefore useful to track as both a model release and a practical reference point for teams evaluating open model workflows.

The model is reported with a native one million token context window. It uses Large-scale Sparse Attention, described as LSA, to support long-context use. The model is therefore relevant for teams that need to review repositories, documents, logs, benchmark notes, and implementation traces together.

The architecture uses a Shortcut-connected MoE approach, described as ScMoE in the release material. It also uses Multi-head Output Proportion Decoupling, described as MOPD. The model is presented as a sparse model that balances activation cost, capacity, and long-context behavior.

The model was trained on more than 30 trillion tokens according to the official release material. Its training is also described as using a 50,000-card domestic compute cluster. The model is therefore discussed not only as a model artifact, but also as an infrastructure milestone.

The release reports 59.5 on SWE-bench Pro. LongCat 2.0 is also reported at 70.8 on Terminal-Bench 2.1. The model should not be reduced to two numbers, but those reported benchmarks explain why coding-agent evaluators are paying attention.

The release reports 78.8 on RWSearch, 73.2 on FORTE, and 79.9 on BrowseComp in official release material. The model is therefore presented with agentic search and browsing benchmarks alongside coding benchmarks. This site keeps those figures separated from independent interpretation.

LongCat 2.0 Online is independent from the official LongCat project. This site links to source material and summarizes public facts for readers. This site does not claim to be the official LongCat 2.0 release page, repository, benchmark owner, or model provider.

The release appears in public discussion because it combines large sparse capacity with a smaller active computation path. It is not the first MoE model, but the reported 1.6T total parameter scale makes LongCat 2.0 important for readers comparing open releases, coding benchmarks, and inference tradeoffs.

Official material emphasizes agentic coding, agentic search, and long-context operation together. This site repeats that grouping because it is central to the public positioning. The model should therefore be compared with models and systems designed for tool use, repository work, and multi-step reasoning.

Benchmark references should be read with care. Reported scores describe performance on named test sets, not every possible software task. This site encourages readers to treat public scores as starting points for evaluation rather than as a substitute for local testing.

Source links matter because model information changes over time. This site uses source links so readers can verify release details, benchmark values, license statements, and community reactions. This site should be updated when official LongCat 2.0 material changes.

This site avoids unsupported claims about access, pricing, hosted inference, or private integrations. This site describes LongCat 2.0 with public facts first. This site can add product features later, but the content baseline should remain tied to verifiable LongCat 2.0 sources.

For search clarity, LongCat 2.0 Online keeps the exact phrase LongCat 2.0 on important factual references. LongCat 2.0 is the model name readers are searching for, and LongCat 2.0 should appear where the page states model size, context length, license, and reported benchmarks.

LongCat 2.0 Online uses the phrase LongCat 2.0 carefully rather than stuffing it into every sentence. LongCat 2.0 still appears often enough for topical relevance, while surrounding sentences use normal references such as the model, the release, the architecture, and this site.

The practical question is how LongCat 2.0 facts help a reader decide what to test. LongCat 2.0 Online answers that question by connecting LongCat 2.0 public data to evaluation planning, source review, and agentic coding workflow design.

LongCat 2.0 is the central entity on this site. LongCat 2.0 facts are repeated where they identify the model, not where ordinary prose can use a pronoun. LongCat 2.0 density is therefore intentional but moderated.

LongCat 2.0 appears in architecture notes, benchmark notes, and license notes. LongCat 2.0 also appears in source-review language because source verification is part of the page purpose. LongCat 2.0 remains the exact keyword for the topic.

LongCat 2.0 Online keeps LongCat 2.0 wording near claims that readers may want to verify. LongCat 2.0 wording is less useful in generic sentences, so those sentences use shorter references. LongCat 2.0 content should read naturally.

LongCat 2.0 is the exact model phrase used for SEO checks. LongCat 2.0 appears here to identify the topic, and LongCat 2.0 remains tied to factual statements.