LongCat-2.0, a large-scale MoE model with 1.6T total and 48B Active

LongCat-2.0 just released a 1.6 trillion parameter MoE model that activates only 48 billion weights per token.

The announcement came from the LongCat team on June 30, 2026. The model targets production inference rather than research benchmarks. Knowledge workers see immediate value in faster synthesis across documents and meetings.

LongCat-2.0 keeps total size high while cutting active compute. This split changes how teams can run large models inside daily tools without new hardware.

Model release details

LongCat-2.0 uses a mixture of experts design. The full network holds 1.6 trillion parameters. Each forward pass routes tokens to a subset that equals 48 billion active parameters.

The team published the numbers on their site. They report training on a mix of public web data and licensed enterprise text. No further training details appeared in the first post.

Inference runs on standard GPU clusters already used by many firms. The active parameter count stayed low enough to fit inside existing context windows for most office tasks.

Knowledge workers testing early access noted stable output across long reports. One legal operations analyst synthesized 40-page cross-border merger agreements, extracting clause inconsistencies and jurisdiction-specific obligations, in under two minutes. The same task took six minutes on the prior dense 70 billion model.

Why the split matters now

Most production agents still trade size against speed. Teams want broader knowledge coverage yet refuse to accept slower replies during meetings or research sprints.

LongCat-2.0 attacks that tradeoff directly. It keeps total capacity high for rare edge cases while routing most tokens through a narrow active path.

The timing aligns with rising demand for local or private inference. Many organizations now run models on their own clusters to avoid sending meeting notes to third party APIs.

This release gives those teams a new option. They can load the full 1.6 trillion parameter checkpoint, then let the router decide which experts fire. Active load stays close to a 48 billion model.

Comparison with prior MoE releases

Earlier MoE systems often required custom routing code or suffered from load imbalance. LongCat-2.0 ships with a simpler dispatch layer that runs on standard frameworks.

The active parameter target of 48 billion sits between common dense models. It beats the 32 billion class on knowledge breadth yet stays faster than 70 billion dense alternatives.

Knowledge workers who compared outputs side by side saw fewer factual slips on technical topics. The larger expert pool appears to reduce hallucination on niche industry terms.

Still, the company does not claim universal superiority. It states only that routing quality improved through targeted training on enterprise style documents.

Limits that remain

The 48 billion active figure holds only when the router behaves as expected. Heavy out of domain queries can push more experts online and raise latency.

Early testers reported occasional spikes when handling mixed language inputs. The team acknowledged this in the release notes and said a follow up patch will tighten the router.

No public benchmark leaderboard entry exists yet. LongCat-2.0 therefore sits outside direct head to head comparisons with other frontier releases.

Knowledge workers should treat the 1.6 trillion total size as a capacity ceiling rather than a daily working size. Real usage stays closer to the active slice.

How teams can test it today

The model weights sit behind a gated download. Approved users receive an inference script that works with common serving libraries.

One firm integrated the router into an internal search tool. Meeting summaries now pull context from years of archived files without manual uploads.

Another group plans to pair the model with local retrieval systems. They keep all documents on site and let the 48 billion active path handle synthesis.

These pilots remain small scale. Larger rollouts will need more data on token routing stability over weeks of continuous use.

Next signals to watch

Watch for the promised router update in the next eight weeks. A drop in out of domain latency would strengthen the production case.

Track third party benchmark submissions. Independent numbers on enterprise document sets would give knowledge workers clearer signals.

Observe whether other labs copy the same total to active ratio. If several releases adopt similar splits, the pattern points to a new default for office grade models.

LongCat-2.0 does not end the size versus speed debate. It simply shows one workable path that fits inside current hardware budgets while expanding the knowledge surface available to each query.