LongCat-2.5 Launches with Native Multimodality and 1.6T Parameters for Agentic Coding

• LongCat-2.5 is a 1.6T parameter Mixture-of-Experts model with 48B active parameters per token, featuring native multimodality and enhanced agentic coding capabilities.
• The model scores 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp, significantly surpassing its predecessor LongCat-2.0 (70.8 and 79.9 respectively) and competing with GPT-5.5 and Claude-Opus-5.
• Visual processing uses Residual Vector Quantization (RVQ) on native resolution features to integrate visual and text tokens into a shared semantic space, eliminating information bottlenecks.
• Agentic Reinforcement Learning (RL) employs asynchronous rollouts, staleness correction via log-prob clipping, and reward hacking prevention through network isolation and LLM-based trajectory auditing.

Technical Architecture

• Multi-Teacher On-Policy Distillation (MOPD) uses full-vocabulary KL divergence from domain experts to guide the student model, with Tree Attention enabling single forward/backward pass efficiency.
• Infrastructure optimized for domestic AI accelerators with under 80GB memory per device, utilizing Supernode-aware parallelism and dynamic expert replication to prevent OOM.
• W4A4 quantization solution improves prefill throughput by 43.3% and decode throughput by 42.77% compared to W8A8 INT8 baselines.

Evaluation Context

• Benchmarks include Terminal Bench v2.1, DeepSWE v1.1, BrowseComp, and MathVision, tested with specific harnesses like Claude Code and mini-swe-agent.
• Comparison targets include GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8, Claude-Opus-5, GLM-5.3, kimi-k3, and DeepSeek-V4.1-Flash.