LongCat-2.0

OpenCode's official documentation for Go confirms LongCat-2.0 is included in the curated lineup of open coding models available through the $10/month subscription. The docs page, published seven hours ago, lists LongCat-2.0 alongside other models including Grok 4.6, GLM-5.x variants, Kimi K3 and K2.7 Code, DeepSeek V4, The docs also note OpenCode Go is designed primarily for international users with stable global access, and traffic is monitored to flag abusive usage. For LongCat-2.0 specifically, this documentation positions it as a first-class citizen in OpenCode's $10/month tier for agentic coding workloads. The model joins offeri

This analytical piece frames Meituan's LongCat-2.0 release as geopolitically significant: a 1.6-trillion-parameter model trained and served entirely on domestic Chinese chips with no Nvidia GPUs or Google TPUs, using 50,000+ Chinese-made accelerators (likely Huawei Atlas-950 SuperPods, though unconfirmed by Meituan). B The piece highlights pricing as the model's competitive lever, with $0.30 per million input tokens against GPT-5.5's $5 and Opus 4.8's $15, making the quality-per-dollar ratio the central decision factor for adopters. It contextualizes the release as a direct empirical challenge to U.S. export-control assumptions about

This Yahoo/Tech report confirms Meituan officially unveiled LongCat-2.0 on June 30, identifying it as the same system that operated anonymously on OpenRouter under the alias "Owl Alpha" for two months. The model is a 1.6-trillion-parameter MoE that activates roughly 48 billion parameters per token, swinging between 33B The report emphasizes that LongCat-2.0 is the first trillion-parameter model trained and deployed end-to-end on domestic Chinese ASICs, with pretraining spanning over 35 trillion tokens across 50,000+ domestically produced accelerators and finishing without rollbacks or irrecoverable loss spikes. API pricing is listed

Yahoo Tech / BeInCrypto reports that Meituan unveiled LongCat-2.0 on June 30, 2026, positioning it as China's biggest AI model trained entirely without NVIDIA chips. The 1.6-trillion-parameter open-source LLM was built end-to-end on domestic hardware, with Meituan claiming it is the first trillion-parameter model to co The compute stack is built around large-scale ASIC superpods coordinated via Huawei's Collective Communication Library (HCCL), mirroring NVIDIA's NCCL role in GPU clusters. LongCat-2.0 supports a 1M-token context window with 48B active parameters per token (per Meituan's launch tweet) and uses LongCat Sparse Attention

OpenRouter's model page independently corroborates LongCat 2.0's core specifications: a sparse mixture-of-experts model from Meituan with 48 billion active parameters out of 1.6 trillion total, suited for coding, repository-level changes, long-horizon problem solving, and agentic workflows. The model supports a 1-milli Performance metrics from AtlasCloud show throughput of 33 tokens per second at P50, latency of 3.33 seconds P50, and a tool call error rate of 2.02%. The cache hit rate averages 85.80% and uptime has been 100% over the past three days. These metrics provide a benchmark for what users might expect when accessing LongCat

LongCat officially released and open-sourced LongCat-2.0, a 1.6-trillion-parameter MoE model purpose-built for agentic coding, according to the LongCat AI News & Updates page. It is described as the industry's first trillion-parameter model to complete full training and inference on a 50,000-card domestic compute clust The release highlights several architecture components: LongCat Sparse Attention (LSA) for linear-complexity sparse attention over 1M tokens, Zero-computation experts combined with ScMoE for token-level dynamic compute allocation, and MOPD multi-expert fusion that dynamically routes between Agent, Reasoning, and Intera