LongCat 2.0

OpenRouter lists Meituan's LongCat 2.0 (model id meituan/longcat-2.0) as a sparse mixture-of-experts language model with 48B active parameters out of 1.6T total, released on OpenRouter on July 20, 2026, and exposed via a single host (AtlasCloud) with no routing across multiple providers. The page documents a native 1-m For developers choosing a model on OpenRouter, the listing frames LongCat 2.0 as suited to coding, repository-level changes, long-horizon problem solving, and agentic workflows, with the 1M context and MoE efficiency positioned as the headline capabilities. The single-provider setup means performance, availability, and

TechJack Solutions reports that Meituan open-sourced LongCat-2.0 on July 6, 2026 under an MIT license, publishing the model on Hugging Face. The piece characterizes the release as a "reveal" rather than a debut, noting that a model matching the description was already serving on OpenRouter under the name "Owl Alpha" be The article reiterates Meituan's headline numbers — 1.6 trillion total parameters with about 48 billion active per token — and surfaces the vendor's self-reported benchmark claims that LongCat-2.0 outperforms Claude Opus on IFEval and IMO-AnswerBench. It explicitly flags that no independent third-party evaluation has b

MarkTechPost's July 5, 2026 launch coverage describes LongCat-2.0 as Meituan's next-generation trillion-parameter open model, following the 2025 LongCat-Flash (560B), and reports a 1.6T-total / ~48B-active MoE architecture with a native 1-million-token context window. The post emphasizes an agentic-coding focus, coveri The same article highlights that both training and serving for LongCat-2.0 were run entirely on domestic AI ASIC superpods rather than on NVIDIA GPU clusters, framing the release as a step in Meituan's vertically integrated infrastructure stack. As a third-party tech blog the piece is useful for architectural and capab

According to a July 3, 2026 technical write-up at The Tech Society, Meituan open-sourced LongCat-2.0 on June 30, 2026 as a 1.6-trillion-parameter sparse mixture-of-experts language model with 48B active parameters and a 1-million-token context window, positioned for coding, repository-level changes, long-horizon proble The same article notes API pricing at $0.30 per million input tokens, undercutting GPT-5.5 ($5) and Opus 4.8 ($15), framing the release as a quality-per-dollar play that directly challenges assumptions behind U.S. chip export controls. Several specific claims in the piece — the anonymous "Owl Alpha" OpenRouter presence

South China Morning Post reported on June 30, 2026 that Meituan released what it claims is China's largest artificial intelligence model trained entirely on home-grown hardware, as Beijing pushes to move beyond using domestic chips solely for inference. LongCat-2.0 boasts 1.6 trillion parameters and a million-token con The article, authored by Xinmei Shen and published at 1:21pm HKT with a 2:59pm update, is a 2-minute read aimed at a general business-tech audience. The supplied excerpt is dominated by site chrome, paywall, and cookie-consent UI rather than technical body text, so benchmark scores and architecture details are not extr

A June 30, 2026 news flash confirms Meituan's release of LongCat-2.0, a 1.6-trillion-parameter AI model with a 1-million-token context window. According to the brief, citing SCMP, the model was trained and deployed entirely on domestic hardware using Huawei's Collective Communication Library (HCCL) for chip-to-chip coo As a third-party pointer sourced from SCMP, the item provides minimal independent technical detail beyond the headline figures — it restates parameter count, context length, HCCL usage, and coding focus without adding benchmark scores or architecture specifics. The page carries a standard financial/investment disclaime

Meituan unveiled LongCat-2.0 on June 30, 2026, debuting what the company calls China's biggest AI model trained entirely without NVIDIA chips. The 1.6-trillion-parameter open-source LLM matches the scale of DeepSeek's flagship with a 1-million-token context window, and was trained end-to-end on domestic hardware — dist The release post on X (embedded in the article) lists 1.6T parameters, MoE with 48B active, 1M-token context, and LongCat Sparse Attention (LSA) for efficient 1M-context scaling, confirming the full model behind the earlier "Owl Alpha" anonymous deployment on OpenRouter. Analyst Yuchen Jin commented that export control

The LongCat-aligned benchmarks page reports author-claimed agentic scores for LongCat-2.0 from the June 2026 release: 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. FORTE ties Claude Opus 4.6 (GPT-5.5 reports 77.8) and SWE-bench Pro edges GPT-5.5 at 59.5 versus 58.6. The page also documents LongCat-Flash text benchmark results, including MMLU 89.71, IFEval 89.65, MATH500 96.40, DROP F1 79.06, Humaneval+ pass@1 88.41, and tau-squared-Bench telecom avg@4 73.68. A new UNO-Bench is introduced as a 1,250-sample omni-modality suite with 2,480 single-modality items and 98% requiring cross-modal fusion, using a 95%-accurate automatic scoring model.