LongCat-2.0

On June 30, 2026, LongCat introduced and open-sourced LongCat-2.0, a large-scale mixture-of-experts language model with 1.6 trillion total parameters and 48 billion activated per token. The full training run and large-scale deployment were built entirely on AI ASIC superpods, with pretraining spanning millions of accel LongCat-2.0 is positioned for long-horizon and agentic workloads, integrating with mainstream harnesses such as Claude Code, OpenClaw, and Hermes for code understanding, repository-level edits, automated task execution, and agentic workflows. The architecture builds on LongCat-Flash with further parameter-efficiency im

LongCat officially released and open-sourced LongCat-2.0, described as a next-generation 1.6 trillion-parameter Mixture-of-Experts model targeted at agentic coding and positioned as the first trillion-parameter model to complete full training and inference on a 50,000-card domestic compute cluster. The page documents n The release page lists vendor-reported benchmarks: SWE-bench Pro 59.5, SWE-bench Multilingual 77.3, Terminal-Bench 2.1 70.8, RWSearch 78.8, FORTE 73.2, and BrowseComp 79.9, and notes that before the official announcement the model operated anonymously as Owl Alpha on OpenRouter for two months, consuming 10.1 trillion t

MarkTechPost reports that Meituan released LongCat-2.0 as a large-scale Mixture-of-Experts language model with 1.6 trillion total parameters and approximately 48 billion active parameters per token, targeting agentic coding tasks such as code understanding, generation, and execution inside agent workflows. The article The MarkTechPost write-up frames LongCat-2.0 as a significant open-weights MoE release, noting that Meituan open-sourced the model rather than releasing it as a closed API-only product. The article echoes the architectural emphasis on a long context window and a relatively small active-parameter count per token, which

An independent long-form analysis dated July 3, 2026 confirms that Meituan open-sourced LongCat-2.0 on June 30, 2026 as a 1.6 trillion-parameter AI model trained and served entirely on domestic Chinese chips, with a 50,000+ accelerator cluster using the Huawei Collective Communication Library (HCCL). The article adds c The piece reports a vendor benchmark of 59.5 on SWE-bench Pro and notes launch pricing of $0.30 per million input tokens, while contextualizing the release as an empirical challenge to U.S. export-control assumptions about frontier-scale training on non-Nvidia hardware. It frames LongCat-2.0's quality as closer to Clau

Yahoo-aggregated coverage dated July 1, 2026 confirms Meituan's official unveiling of LongCat-2.0 on June 30 as the same system that operated anonymously on OpenRouter under the Owl Alpha alias for two months. The article specifies an average of roughly 48 billion parameters activated per token, swinging between 33B an The coverage details the training run as exceeding 35 trillion tokens across a cluster of over 50,000 domestically produced accelerators, with Meituan claiming the pretraining finished with "no rollbacks or irrecoverable loss spikes" — described as the first trillion-parameter model trained and deployed end-to-end on d

A short news flash dated June 30, 2026, attributing its information to SCMP, confirms that Meituan released LongCat-2.0, a 1.6-trillion-parameter AI model with a 1 million-token context window, on that date. The item specifies that the model was trained and deployed entirely on domestic hardware using the Huawei Collec The flash notes that LongCat-2.0 matches DeepSeek-V4-Pro in context window length and parameter scale, with a stated focus on coding applications. The article carries a disclaimer that the content is third-party reference material and does not represent the views of the exchange, with no additional benchmarks, architec

OpenRouter lists LongCat 2.0 under the provider key meituan/longcat-2.0 as a Mixture-of-Experts language model with 48 billion active parameters out of 1.6 trillion total, suited to coding, repository-level changes, long-horizon problem solving, and agentic workflows. The listing shows a 1-million-token context window Performance telemetry from OpenRouter over the prior week shows P50 throughput averaging 26 tokens per second, P50 end-to-end latency of 13.51 seconds, and 100% uptime for the AtlasCloud deployment. AtlasCloud is the sole host carrying the full token share, meaning there is no provider redundancy on the marketplace tod