LongCat-2.5 Boosts Agentic Reasoning and Coding
LongCat presents LongCat-2.5, a multimodal mixture-of-experts model announced with 1.6 trillion parameters. The publisher highlights improved results in coding, web navigation, and agentic tasks, with an API already open.
LongCat unveiled LongCat-2.5 on November 1, 2026, a native multimodal mixture-of-experts model designed for long agentic tasks, code, and vision. The publisher claims a total of 1.6 trillion parameters and about 48 billion activated per token, with an API already available but GitHub and Hugging Face repositories still announced as coming soon. The launch matters because it combines a context of one million tokens, integration with agent tools, and a stated ambition to accelerate large software projects.
A model targeting long tasks, not just quick answers
LongCat-2.5 is based on the idea that a useful model in business must go the distance. The publisher says it has enhanced native visual understanding and agentic coding, meaning the ability to chain actions, read tools, correct, rerun, and complete a complex task without losing track at the first detour. In this logic, training includes more long trajectories and large-scale reinforcement for developing complex codebases and end-to-end project builds.
The visual part follows the same unification logic. LongCat-2.5 represents visual signals as multi-level discrete tokens to process text and image in a common format. The system combines native resolution visual tokenization, offline preprocessing, and embeddings learned jointly with the backbone. All relies on residual vector quantization applied to features extracted by a semantic and aligned encoder.
Why code and vision were put in the same basket
The choice is not trivial. In practice, LongCat says it can precompute and cache visual token IDs, making data throughput more predictable during training. According to the company, this setup also helps maintain an approximately uniform distribution of modalities across different compute ranks. On a small MoE of about 3 billion parameters activated on average, preliminary validation reportedly showed, via t-SNE visualizations, an interleaving of visual and textual embeddings in a shared semantic space.
For agentic code tasks, LongCat describes a synthesis pipeline more concrete than glamorous. The system filters GitHub pull requests and related issues, has agents annotate tasks, generates interactive environments, and verifies them with fail-to-pass and pass-to-pass tests. Real-world failure trajectories are then converted into reproducible training tasks. In other words, at LongCat, bugs finally serve a purpose. The company adds it disabled network access in environments, removed Git history and cached files, then used an LLM judge to spot exploits and limit reward hacking, when a model learns to cheat the reward rather than solve the problem.
An architecture designed to keep pace
The infrastructure aspect is just as central. LongCat indicates it trained long sequences on domestic AI accelerators with less than 80 GB of memory per device, relying on supernode topology to keep heavy exchanges within a high-bandwidth interconnect. The dual-shard FSDP method splits model weights inside each supernode, then further distributes FP32 master weights and optimizer state across FSDP domains.
The company also claims to have achieved 86% more throughput per device than its DDP baseline on long sequence supervision tasks. Similarly, LongCat says it preserves bitwise identical log probabilities per token between training and inference, with near-zero performance cost, while separating parallelism configurations. A technical promise mainly interesting for teams wanting to run the same system in training and production without surprises.
Improved results on paper, with cautious reading
Scores published by LongCat show a clear jump over LongCat-2.0 on several benchmarks. Terminal Bench v2.1 rises from 70.8 to 88.4, DeepSWE from 9.7 to 64.6, BrowseComp from 79.9 to 91.9, and AutomationBench from 15.8 to 51.8. On MCP Atlas, the score reaches 83.1 versus 61.8, and on HLE with tools, 63.1 versus 37.5. For visual tasks, LongCat announces 55.4 on 3DCodeBench Text-to-3D and 78.6 on Vision2Web.
These figures deserve careful interpretation. The company specifies that some values in its full table come from technical reports or public leaderboards, while others are self-evaluated and merely indicative. Tests were conducted with chain-of-thought mode enabled, a temperature of 1.0, and a top_p of 0.95, which matters for score interpretation. Benchmarks never tell the whole story, but here they give a clear signal: LongCat wants to be judged on its ability to chain complex tasks, not just to chat nicely.
API, pricing, and access: what is already available
For now, public access is mainly through the chat interface and API. The quick start guide lists LongCat-2.5 in OpenAI and Anthropic compatible formats, with endpoints at api.longcat.chat/openai and api.longcat.chat/anthropic. A call can use a context window of 1 million tokens and produce up to 128,000 tokens in output.
The product page also says it supports Claude Code, Codex, Hermes, and other agent harnesses, with announced uses in visual coding, creative design, 3D tasks, complex projects, and multi-step reasoning. On pricing, LongCat temporarily shows $0.30 per million uncached input tokens, $0.006 per cached input token, and $1.20 per output token, respectively 2 yuan, 0.04 yuan, and 8 yuan. Prices may change, and the platform specifies that displayed prices and billing statements are authoritative.
By the numbers
- 1.6 trillion parameters — total announced for LongCat-2.5.
- 48 billion parameters — activated per token, according to LongCat.
- 1 million tokens — maximum context per request, via the API.
- 86% — throughput gain per device compared to the DDP baseline.
- 43.3% — increase in prefill throughput compared to an INT8 W8A8 baseline.