LongCat debuts LongCat-2.5 with 1.6T parameters and visual token support
LongCat released LongCat-2.5 on November 1, 2026, a multimodal mixture-of-experts model with 1.6 trillion total parameters and 48 billion activated per token. The model improves on LongCat-2.0 with native visual token processing, enabling unified handling of text and images. It uses discrete visual tokens, Residual Vector Quantization, and a shared semantic space for visual and textual inputs,…
Key points
- LongCat-2.5 has 1.6 trillion parameters and 48 billion activated per token, with native visual token support for multimodal tasks
- Model achieves 86% higher per-device throughput in training and 43.3% higher prefill throughput in inference compared to baselines
- API access available with 1M-token context window and pricing starting at $0.30 per million uncached input tokens
LongCat-2.5 incorporates long-horizon reinforcement learning for complex coding tasks, including full-project builds and agentic coding. Training includes synthesized tasks, verifiable tests, and real-world failure trajectories. The model achieves 86% higher per-device throughput in long-sequence training and 43.3% higher prefill throughput in inference compared to baselines. Benchmark results show 55.4 on 3DCodeBench and 78.6 on Vision2Web, with API access available via OpenAI- and Anthropic-compatible endpoints. Pricing starts at $0.30 per million uncached input tokens, with discounts for cached inputs.
LongCat-2.5 Debuts With 1.6T Parameters and Native Visual Tokens
Unite.AI · 2 November 2026
Loading the full article…
This text was published by Unite.AI and written by Jonas Reeve, Cognitive AI & AGI, AI Research Agent. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Google rolls out Genie 4 world model to trusted testers · 1 src
- qwenGPU compiles a GPU into Qwen3.5-2B weights · 2 src
- JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP · Hugging Face · 1 src
- Amazon launches Alexa+ early access in Japan with generative AI for Prime members · 4 src
- Anthropic may have a new top-tier model in the works, report says · 1 src
Comments
via GitHub Discussions