LongCat introduced its LongCat-2.5 model on November 1, 2026. The model uses a multimodal mixture-of-experts architecture designed for long-running agent tasks. It has 1.6 trillion total parameters, but only about 48 billion of them are activated per token. The company says the new version adds visual understanding and agent-based coding capabilities on top of LongCat-2.0.
The model is currently accessible through the chat interface and API. LongCat’s GitHub and Hugging Face links are still marked as “coming soon.”
It also processes visual data as tokens
LongCat-2.5 converts images into multi-level discrete tokens. This lets the model backbone process both visual and textual input in a single token format. The system combines natural-resolution visual tokenization, offline visual preprocessing, and visual embedding layers learned together with the backbone. Image features are converted into a token sequence with Residual Vector Quantization, then reconstructed by a visual decoder.
Because a fixed tokenizer is used, visual token IDs can be precomputed and cached. This makes the visual and text load predictable during data packing; load distribution can be kept balanced across parallel processing units. LongCat reports that in a small trial model with about 3 billion active parameters, they observed visual and textual embeddings merging in a shared semantic space.
Synthetic task pipeline for coding agents
The company set up an automated data generation pipeline for agent-based coding. This pipeline filters GitHub pull requests and linked issue records, labels tasks with agents, sets up interactive environments, and validates with tests. Success rates are used in relabeling and filtering tasks. Offline analysis of false positive and false negative traces also enables tests to be corrected. Agent traces that fail in the real world are turned into reproducible training tasks.
Training uses asynchronous partial rollout. To limit reward hacking attempts, network access is disabled in task environments, Git history and cache files are deleted, and traces are additionally scanned by a language model auditor. Repetitive reasoning and calling the same tool over and over are also penalized.
Benchmark results and 4-bit inference
According to the results LongCat shared, the model rises from 70.8 to 88.4 on Terminal Bench v2.1 and from 9.7 to 64.6 on DeepSWE v1.1. BrowseComp goes from 79.9 to 91.9, MCP Atlas from 61.8 to 83.1, and the tool-using HLE test from 37.5 to 63.1. The AutomationBench score has also increased from 15.8 to 51.8. On the visual side, 55.4 is reported on the Text-to-3D test and 78.6 on the Vision2Web test.
On the inference side, expert weights and activations are reduced to 4 bits and run on INT4 units. LongCat says this compression gives accuracy close to the uncompressed model on the evaluated tasks. In stress tests the company measured against an INT8 reference, prefill speed increased by 43.3 percent and decode speed by 42.8 percent; time to first token was shortened by 15.9 percent and time per token by 19.7 percent.
On the API side, there are both OpenAI- and Anthropic-compatible endpoints. A single request supports 1 million token context and 128 thousand token output. Introductory prices are listed as $0.30 per million tokens for uncached input, $0.006 for cached input, and $1.20 for output. LongCat states that the model is compatible with agent tools such as Claude Code, Codex, and Hermes.
Source: Unite.AI
Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.