LongCat-2.5 Debuts With 1.6T Parameters and Native Visual Tokens

LongCat introduced LongCat-2.5 on November 1, 2026, a natively multimodal mixture-of-experts model for long-horizon agentic tasks with 1.6 trillion total parameters and approximately 48 billion activated per token.

LongCat said the release builds on LongCat-2.0 with stronger native visual understanding and agentic coding. Its training incorporates more long-horizon trajectories and large-scale long-horizon reinforcement learning for complex codebase development and full-project builds. The release links to LongCat’s chat interface and API platform, while its GitHub and Hugging Face buttons are labelled as coming soon.

Discrete Visual Tokens and Agent Training

LongCat-2.5 represents visual signals as multi-level discrete tokens, allowing its backbone to process visual and textual inputs in a unified token-based format. The design combines native-resolution visual tokenization, offline visual preprocessing and visual embeddings learned jointly with the backbone. It applies Residual Vector Quantization to native-resolution features from a Semantic-and-Aligned Encoder, with images encoded as discrete token sequences and reconstructed by a visual de-tokenizer.

A fixed tokenizer allows visual token IDs to be precomputed and cached offline, so the backbone processes pre-tokenized sequences. Because image-token counts are known during data packing and shuffling, modality loads are predictable and their proportions can be kept approximately uniform across data-parallel ranks. LongCat said preliminary validation on a smaller MoE with approximately 3 billion activated parameters on average showed interwoven visual and textual embeddings in t-SNE visualizations, suggesting integration within a shared semantic space.

For agentic coding tasks, a synthesis pipeline generates training tasks, interactive environments and verifiable tests. LongCat filters GitHub pull requests and linked issues, uses agents to annotate tasks and synthesize environments, and validates them with fail-to-pass and pass-to-pass tests. Live rollout pass rates guide relabeling and filtering, while offline analysis of false-positive and false-negative trajectories corrects and extends the tests. Real-world agent failure trajectories are also converted into reproducible, verifiable training tasks.

Asynchronous partial rollouts interrupt and save in-flight trajectories after a full batch is collected, then resume them during the next sampling phase. To limit reward hacking, LongCat disables network access in task environments, removes Git history and cached files, and uses an LLM judge to audit trajectories for exploits. Additional penalties address repetitive reasoning and repeated tool calls during multi-turn interactions.

Accelerator Training and W4A4 Inference

Long-sequence training occurs during both mid-training and supervised fine-tuning on domestic AI accelerators with less than 80 GB of memory per device. Parallel groups follow the supernode topology, keeping parameter all-gather, expert token all-to-all and KV all-gather operations within its high-bandwidth interconnect.

Dual-shard FSDP keeps model-parameter sharding within each supernode while further sharding FP32 master weights and optimizer states across FSDP domains. LongCat reported 86% higher per-device throughput than its DDP baseline in long-sequence supervised fine-tuning tasks.

LongCat said its training and inference stack preserves bitwise-identical per-token log probabilities with near-zero performance overhead while decoupling their parallelism configurations. Both engines use batch-invariant kernels and aligned framework logic, with backward support for training.

W4A4 inference quantizes expert weights and activations to 4 bits and executes the corresponding matrix multiplications on native INT4 units. A shared input transformation across experts enables quantization before token replication and dispatch, allowing the same INT4 representation to be reused during expert communication and computation. LongCat reported accuracy comparable to the unquantized model on all evaluated downstream tasks.

In online stress tests against a W8A8 INT8 baseline, LongCat reported 43.3% higher prefill throughput and 42.77% higher decode throughput, alongside reductions of 15.91% in TTFT and 19.72% in TPOT. The GMM kernel achieved an average 1.73× speedup over W8A8.

Reported Evaluation Results

LongCat reported the following benchmark results for LongCat-2.5 and its predecessor:

For visual evaluations, LongCat reported 55.4 on 3DCodeBench Text-to-3D and 78.6 on Vision2Web.

Unless otherwise specified, the LongCat-2.5 evaluations used thinking enabled, a temperature of 1.0 and top_p of 0.95. Terminal Bench v2.1 used the Claude Code harness with a 512K maximum sequence length and was averaged over three independent runs. DeepSWE v1.1 used mini-swe-agent with a 1M maximum sequence length and was also averaged over three runs.

For other models included in LongCat’s full evaluation table, scores marked with an asterisk came from public technical reports, blog posts or official benchmark leaderboards. Unmarked scores were self-run and described as indicative only.

API Access and Pricing

The API quick-start guide lists LongCat-2.5 for both OpenAI- and Anthropic-compatible formats. Their respective base endpoints are https://api.longcat.chat/openai and https://api.longcat.chat/anthropic, supporting /v1/chat/completions and /v1/messages. A single request has a 1M-token context window and a maximum output length of 128K tokens.

The LongCat product page lists compatibility with Claude Code, Codex, Hermes and other agent harnesses. It also describes support for visual coding, creative design and 3D tasks, alongside complex codebases and full-project builds involving multi-step reasoning and tool use.

Pay-as-you-go pricing lists limited-time discounted rates per million tokens of $0.30 for uncached input, $0.006 for cached input and $1.20 for output. The corresponding yuan rates are ¥2, ¥0.04 and ¥8. Prices may change, with the prices displayed on the API platform and billing records treated as authoritative.

With GitHub and Hugging Face still marked as coming soon, the release presents its chat interface and API as the available access routes.