Meituan has launched LongCat-2.0, a large-scale Combination-of-Consultants (MoE) language mannequin. It carries 1.6 trillion whole parameters and prompts about 48 billion per token. The mannequin targets agentic coding: code understanding, era, and execution inside agent workflows.
Two info stand out. First, LongCat-2.0 helps a local 1-million-token context window. Second, each coaching and serving ran fully on home AI ASIC superpods.
What’s LongCat-2.0?
LongCat-2.0 is Meituan’s next-generation trillion-parameter open mannequin. It follows LongCat-Flash, a 560B mannequin launched in 2025. The structure was designed round one objective: dependable, environment friendly agentic coding.
Pretraining spanned greater than 35 trillion tokens over thousands and thousands of accelerator-hours. Meituan experiences no rollbacks or irrecoverable loss spikes throughout the run. That stability declare issues on non-Nvidia {hardware}, the place tooling is much less mature.
Structure: How a 1.6T Mannequin Stays Low-cost to Run
The design combines 4 concepts that cut back the price of scale. Every one is price understanding by itself.
- Zero-computation consultants: Not each token wants heavy compute. Easy tokens like punctuation path to a zero-computation knowledgeable and return unchanged. Advanced tokens interact extra knowledgeable capability. A PID controller adjusts knowledgeable bias to carry the typical in vary. This produces the 33B–56B dynamic activation window as a substitute of a hard and fast value. The MoE spine makes use of a shortcut-connected design (ScMoE) for greater throughput.
- LongCat Sparse Consideration (LSA): Customary consideration scales quadratically with context size. LSA selects solely essentially the most related tokens, dropping the scaling nearer to linear. Meituan describes it as an evolution of DeepSeek Sparse Consideration (DSA). It layers three orthogonal indexing strategies. Streaming-aware Indexing turns fragmented reminiscence reads into contiguous blocks. Cross-Layer Indexing reuses consideration saliency throughout adjoining layers. Hierarchical Indexing applies coarse-to-fine two-stage filtering. Collectively they maintain the 1M-token window and not using a reminiscence wall.
- N-gram Embedding: The design provides a 135-billion-parameter N-gram embedding module. It sits orthogonal to the MoE consultants in sparse dimensions. Meituan says it captures dense native token relationships. It additionally reduces reminiscence I/O throughout large-batch decoding.
- Put up-training (MOPD): A devoted pipeline (MOPD) fuses three trainer knowledgeable teams. These cowl Agent, Reasoning, and Interplay capabilities into one unified mannequin.
For serving, Meituan makes use of a 6D parallelism scheme and a prefill-decode disaggregated structure. It additionally employs ‘tremendous kernels’ and L2-cache weight prefetching to cover I/O latency.
Benchmarks
Meituan positions LongCat-2.0 as an agentic coding mannequin. Each determine under comes from Meituan’s personal testing.
| Benchmark | LongCat-2.0 | What it measures |
|---|---|---|
| SWE-bench Professional | 59.5 | Actual-world software program engineering duties |
| Terminal-Bench 2.1 | 70.8 | Execution and error restoration in shells |
| SWE-bench Multilingual | 77.3 | Cross-language repository duties |
On SWE-bench Professional, Meituan experiences LongCat-2.0 edging GPT-5.5 (58.6). Meituan additionally claims general efficiency corresponding to Google’s Gemini 3.1 Professional. The reported edge is concentrated in software program engineering. On broader general-agent benchmarks equivalent to FORTE and BrowseComp, protection signifies it trails main frontier programs. Impartial leaderboard affirmation isn’t but accessible.
LongCat-2.0 vs LongCat-Flash
The bounce from the earlier era is massive on paper. This desk makes use of every mannequin’s printed specs.
| Attribute | LongCat-2.0 | LongCat-Flash |
|---|---|---|
| Whole parameters | 1.6T | 560B |
| Energetic per token | ~48B (33B–56B) | ~27B (18.6B–31.3B) |
| Context window | 1M tokens (native) | 128K tokens |
| Lengthy-context consideration | LongCat Sparse Consideration | Multi-head Latent Consideration |
| Reported {hardware} | Home AI ASIC superpods (coaching + serving) | H800 GPUs (inference reported) |
| Max output | 128K tokens | Not specified |
| License | MIT | MIT |
| Launched | June 30, 2026 | September 2025 |
| Weights | Coming quickly | Open |
Use Circumstances With Examples
LongCat-2.0 is tuned for agent-style software program work, not informal chat. A number of concrete patterns match its strengths.
- Entire-repository reasoning: Feed a whole mid-sized codebase into the 1M-token window. Ask the mannequin to hint a bug throughout many recordsdata directly. This avoids the summarization hacks that shorter home windows pressure.
- Multi-step terminal duties: Run the mannequin inside an agent loop with shell entry. It may execute instructions, learn errors, and retry till a job passes. The Terminal-Bench 2.1 focus targets precisely this workflow.
- Repository-level edits: Ask for a refactor that spans a number of modules and assessments. The mannequin causes over the total context earlier than proposing coordinated adjustments.
- Cross-language migration: Use the SWE-bench Multilingual energy for polyglot repositories. The mannequin can port logic between languages whereas preserving conduct.
These patterns run inside customary agent harnesses. Dev groups can due to this fact undertake the mannequin with out constructing new tooling.
The way to Entry It
LongCat-2.0 is reachable by way of the LongCat API Platform. It exposes each OpenAI-compatible and Anthropic-compatible endpoints. The mannequin can be on OpenRouter and in harnesses like Claude Code, OpenClaw, OpenCode, and Codex. Native self-hosting isn’t but attainable, since weights stay pending.
The OpenAI-compatible endpoint makes use of the mannequin ID LongCat-2.0. Most output size is 131072 tokens (128K). The snippet under calls the documented chat-completions endpoint.
# pip set up openai
from openai import OpenAI
consumer = OpenAI(
api_key="YOUR_LONGCAT_API_KEY",
base_url="https://api.longcat.chat/openai/v1",
)
resp = consumer.chat.completions.create(
mannequin="LongCat-2.0",
messages=[
{"role": "system", "content": "You are a coding agent."},
{"role": "user", "content": "Refactor utils.py to remove duplicate I/O logic."},
],
max_tokens=4096, # LongCat-2.0 helps as much as 131072 (128K)
)
print(resp.selections[0].message.content material)Pricing is reported at $0.75 per million enter tokens and $2.95 per million output. A launch promotion lists $0.30 and $1.20, with cached context reads free. These figures come from third-party protection and should change.
Interactive Explainer
Key Takeaways
- Launched beneath MIT
- LongCat-2.0 is a 1.6T-parameter MoE mannequin activating ~48B parameters per token (dynamic vary 33B–56B).
- Native 1M-token context comes from LongCat Sparse Consideration, reducing long-context value from quadratic to linear.
- Coaching and inference ran on a 50,000-card home AI ASIC cluster, with no Nvidia {hardware}.
- Vendor-reported scores: 59.5 SWE-bench Professional, 70.8 Terminal-Bench 2.1, 77.3 SWE-bench Multilingual.
Take a look at the Mannequin Weights, GitHub Repo and Technical particulars. Additionally, be at liberty to observe us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our E-newsletter. Wait! are you on telegram? now you’ll be able to be a part of us on telegram as properly.
Must associate with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and many others.? Join with us
