LongCat

LongCat officially released and open-sourced LongCat-2.0, a next-generation 1.6-trillion-parameter mixture-of-experts model purpose-built for agentic coding, with native 1M-token context support and dynamic activation of 33B–56B parameters per token. The provider describes it as the industry's first trillion-parameter The same announcement lists provider-reported benchmark scores — 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 — positioning the model against agentic-coding competition. The page also documents the open-source release of LongCat-Video-Avatar 1.

Meituan's LongCat team released LongCat 2.0 on June 30, 2026 as the successor to the LongCat-Flash line, retiring six Flash models a month earlier. Per the model card, it is a 1.6-trillion-parameter mixture-of-experts model with roughly 48 billion parameters activated per token, plus 135 billion N-gram Embedding parame The headline architectural addition is LongCat Sparse Attention (LSA), presented as a fix for two named issues in DeepSeek V3.2's Lightning Indexer — output discontinuity and a quadratic scoring bottleneck. LSA combines Streaming-aware Indexing, Cross-Layer Indexing, and Hierarchical Indexing, though the launch post qu

GIGAZINE's launch coverage confirms LongCat-2.0's core specifications: a 1.6-trillion-parameter MoE model with about 48 billion parameters activated per token and a 1-million-token context window, released by Meituan on June 30, 2026. The article notes the flagship builds on the underlying LongCat-Flash architecture, w Meituan developed LongCat Sparse Attention (LSA) as an evolution of DeepSeek Sparse Attention to address attention bottlenecks for long inputs driven by AI agent workloads, accelerating long-context processing without compromising model quality. A distinctive infrastructure claim is that the full pipeline — pretraining

Reuters, republished on Yahoo Finance on 30 June 2026, confirms that Meituan released and open-sourced LongCat-2.0 as what the company claims is the world's first trillion-parameter AI system trained and run entirely on a 50,000-chip cluster powered by Chinese-made processors, with support for inputs of up to one milli For product impact, the wire-service report highlights that although Meituan did not disclose immediate integration plans, the model is targeted at agentic coding and long-context workflows, and the official statement showcased its ability to build a gaming website and write a novel. The story establishes the geopoliti

Meituan's LongCat large-model team open-sourced LongCat-Video-Avatar 1.5, a commercial-grade digital-human video generation model positioned as moving from open-source SOTA to commercial application readiness, with improvements in lip synchronization, physical plausibility, long-video stability, multi-person interactio The first upgrade swaps the audio feature extraction encoder from Wav2Vec2 to Whisper-large, capturing phoneme and rhythm detail for better lip-sync under long sentences, fast speech, and singing, while reducing frame skipping and identity drift over long videos. The second upgrade introduces a multi-stage data pipelin

AIHubMix's reseller listing documents the open-source LongCat-Flash-Chat model from Meituan as a 560-billion-parameter MoE with a zero-computation expert mechanism that activates roughly 27B parameters per token, and provides concrete API pricing for routing access to it: $0.14 per million input tokens and $0.70 per mi A separate "Longcat 2.0" entry on the same reseller page lists pricing of approximately $0.7746 per million input tokens and $3.0984 per million output tokens, with a measured time-to-first-token of 6.072 seconds and throughput of 38 tokens per second, and describes the tier as designed for agent development with nativ