Meituan LongCat-2.5 — 1.6T MoE for Long-Horizon Agents, Claimed to Approach Claude Opus 5; Weights Not Yet Opened
A MoE with 1.6 trillion total parameters, ~48 billion active per token, natively multimodal (including visual coding/3D), focused on long-horizon agents (large codebases, complete software projects). Meituan's benchmark table: 88.4 Terminal Bench v2.1 vs 89.1 Claude Opus 5 and 88.8 GPT-5.6-Sol — a claimed positi…
A MoE with 1.6 trillion total parameters, ~48 billion active per token, natively multimodal (including visual coding/3D), focused on long-horizon agents (large codebases, complete software projects).
Meituan's benchmark table: 88.4 Terminal Bench v2.1 vs 89.1 Claude Opus 5 and 88.8 GPT-5.6-Sol — a claimed position "between Opus 4.8 and Opus 5."
Process claims: trained entirely on domestic Chinese compute (accelerators under 80GB).
An interesting detail for practitioners: reward-hacking mitigation in its agentic RL — network access is disabled in the task environment, git history and caches are wiped, and an LLM judge audits trajectories.
**Weights not yet released** (GitHub/…