LongCat-2.0: Meituan Open-Sources 1.6T Agent Model with 1M Context — Trained Entirely on Chinese Chips

By Prahlad Menon Published 2026-07-05 3 min read

The mystery model that topped OpenRouter’s developer rankings as “Owl Alpha” has been unveiled. LongCat-2.0 from Meituan is now fully open-source under the MIT license — and it’s a big deal.

The Numbers

Spec	Value
Total Parameters	1.6 trillion
Active per Token	~48B (dynamic 33B-56B)
Context Window	1 million tokens (native)
License	MIT
Training Hardware	50,000+ Chinese AI ASIC chips
Training Data	35+ trillion tokens

This is the first trillion-parameter model to complete full training and inference without relying on NVIDIA GPUs. The entire stack runs on domestic Chinese accelerators.

Agent-Native by Design

LongCat-2.0 isn’t just another LLM — it’s built specifically for agentic coding. It integrates directly with:

Claude Code
OpenClaw
Hermes Agent

The model excels at code understanding, repository-level edits, automated task execution, and complex agentic workflows.

Technical Innovations

LongCat Sparse Attention (LSA)

The 1M context window isn’t just marketing. LSA addresses the quadratic scaling problem with three innovations:

Streaming-aware Indexing (SI) — Reshapes token selection for hardware-aligned sequential reads instead of fragmented memory access
Cross-Layer Indexing (CLI) — Amortizes indexing cost across adjacent layers, since attention saliency stays stable
Hierarchical Indexing (HI) — Two-stage coarse-to-fine scoring shrinks the candidate space per query

Zero-Compute Experts

Dynamic activation ensures simpler tasks consume less computational overhead. The model activates between 33B and 56B parameters depending on task complexity.

N-gram Embedding

135B additional parameters from n-gram embeddings improve parameter utilization efficiency — expanding along sparse dimensions orthogonal to MoE.

Benchmark Performance

LongCat-2.0 competes with frontier models on agentic tasks:

Benchmark	LongCat-2.0	GPT-5.5	Claude Opus 4.7
Terminal-Bench 2.1	70.8	73.8	71.7
SWE-bench Pro	59.5	58.6	64.3
SWE-bench Multilingual	77.3	—	80.5
BrowseComp	79.9	84.4	79.3
FORTE	73.2	77.8	77.6

Note: LongCat-2.0 beats GPT-5.5 on SWE-bench Pro (59.5 vs 58.6).

Deployment

Runs on both GPU and NPU platforms. Recommended GPU setup: 16x H20 with Tensor + Expert Parallelism.

SGLang support is available via PR #30042.

python -m sglang.launch_server \
  --model meituan-longcat/LongCat-2.0-FP8 \
  --trust-remote-code \
  --tp 16 --ep 16 \
  --max-running-requests 64

Why This Matters

Three things stand out:

MIT License — Full commercial use, no restrictions. Modify, redistribute, integrate into proprietary software.
Agent-First Design — Not an afterthought. Deep integration with the tools developers actually use.
Hardware Independence — Proof that frontier-scale training is possible without NVIDIA. Millions of accelerator-days, no rollbacks, no irrecoverable loss spikes.

The open-source AI ecosystem just got a serious new player in the agentic coding space.