Meituan LongCat releases flagship model LongCat-2.0

Meituan LongCat released LongCat-2.0 this week. The model uses a 1.6T parameter MoE design with roughly 48B active parameters. It ships with a native 1M token context window.

The launch places fresh pressure on existing coding agents. Earlier systems often capped context or required heavy compute for large repositories. LongCat-2.0 targets those limits directly.

Model architecture targets long tasks

LongCat-2.0 applies three linked techniques. LSA sparse attention extends the 1M window without quadratic cost growth. Zero-Compute Experts keep activation between 33B and 56B parameters per token. MOPD routes tasks across separate Agent, Reasoning, and Interaction expert groups.

These choices reduce wasted cycles on irrelevant experts. They also allow the model to maintain performance across hour-long coding sessions. SiliconFlow added the model to its platform on day zero.

Pricing and access details

The company listed three price points. Cached input tokens cost 0.015 dollars per million. Standard input tokens cost 0.75 dollars per million. Output tokens cost 2.95 dollars per million.

Developers can test the model through the SiliconFlow platform API today. No waitlist or separate registration is required. The pricing sits between some open models and the highest closed agent services. See Meituan’s official technical release notes for LongCat-2.0.

Performance on real coding benchmarks

LongCat-2.0 scored 59.5 on SWE-bench Pro, a benchmark that tests full-repository issue resolution on real GitHub tasks rather than synthetic snippets. That figure places it near several closed frontier systems. The score reflects full repository edits rather than single-file patches.

The result stems from the task-specific expert routing. Agent experts handle planning. Reasoning experts manage dependency tracing. Interaction experts format patches and tests. Separate routing reduces cross-talk between skills.

Comparison with earlier coding agents

Most prior open models relied on dense attention or shorter windows. Those limits forced developers to split large codebases into chunks. LongCat-2.0 keeps the full context in one pass. A concrete example is refactoring a multi-million-line monorepo containing interdependent microservices: the 1M window lets the model trace cross-file impacts in a single session, whereas earlier models required repeated context resets.

Competing systems often activate every parameter on every token. LongCat-2.0 activates only the needed experts. The difference appears most clearly on multi-hour debugging tasks that span dozens of files.

Remaining open questions

The public release includes limited third-party verification. Independent audits of the 1M context claims have not yet appeared. Long-term stability under heavy agent loops also remains untested at scale.

Meituan has not disclosed training data sources or safety filters. Some observers want clarity on how the model behaves when asked to edit security-critical code. Those details will likely surface only after wider adoption.

What to watch in the next quarter

Watch for updated SWE-bench Pro numbers after community fine-tunes appear. Track whether other platforms match the listed cache pricing. Note any new agent frameworks that ship explicit support for the MOPD routing scheme.

If retention of 1M context proves consistent, more teams may shift from retrieval-augmented setups to single-session agents. If hidden latency costs surface, adoption may stay limited to high-budget users.

Developers focused on large-scale codebases now have another concrete option to test. The model is live and measurable today.