Meituan’s LongCat-2.5 aims at long coding jobs and 3D work

Meituan’s LongCat team says its 1.6 trillion-parameter model handles long coding jobs and 3D tasks. AWS lets agents call Step Functions workflows.

The AI team at Chinese delivery company Meituan has released LongCat-2.5, a model built for long, multistep agent work such as building whole software projects. The model has 1.6 trillion parameters, with about 48 billion used for each token, the team says in a blog post dated Nov. 1.

LongCat says the model reads images and text in one format. It turns pictures into discrete tokens, the same building blocks it uses for words. That pairing of vision and coding supports “creative design and 3D tasks,” the post says.

How it was trained

The team says it trained the model with large-scale reinforcement learning on long jobs, including complex codebase development and full-project builds. It says it turns real cases where agents failed into repeatable training tasks.

To stop the model from gaming its tests, the post says, the team cut off network access in training environments, removed git history and used a second model to audit runs for exploits.

The post says training and inference ran on “domestic AI accelerators” with less than 80 gigabytes of memory each. It does not name the chipmaker. Running weights and activations at 4 bits raised prefill throughput by 43.3% and decode throughput by 42.77% against an 8-bit baseline in stress tests, LongCat says.

The scores

The benchmarks are LongCat’s own. On Terminal Bench 2.1, a test of command-line tasks, it posts 88.4, against 89.1 for Anthropic’s Claude Opus 5 and 90.6 for DeepSeek V4.1 Flash. On the harder Terminal Bench 4.0, it scores 38.5 to Opus 5’s 53.9. On a text-to-3D coding test it scores 55.4, the highest in that row. The post says scores it ran itself for other models are “indicative only.”

The post does not say whether the model’s weights will be released, under what license or how developers can reach it. Its demo videos are marked “coming soon.”

Our read: LongCat’s table is unusually candid. It shows the model trailing rivals on several hard coding tests, not just the ones it wins, and it flags its own runs of other models as indicative. That is more useful than a chart of victories. But every number here comes from the company that built the model, and the post gives no way to try it. A 1.6 trillion-parameter model trained on what the post calls domestic accelerators would be a meaningful engineering feat given U.S. export limits on advanced chips, yet the post does not say which chips or how many. The safeguards against test-gaming are the right ones to describe. Whether they held will show only when outside developers can run the model on their own work. Until then, LongCat-2.5 is a well-documented claim, not a proven tool.

AWS turns workflows into agent tools

Amazon Web Services says workflows built in its Step Functions service can now be called by AI agents as tools, through Amazon Bedrock AgentCore Gateway, according to its weekly roundup Monday. The agent decides when to run a workflow and what to send it; Step Functions runs each step as written. Existing Standard and Express workflows qualify, and agents can look up runs. It costs standard rates wherever the gateway runs, AWS says.

SenseTime posts a looped model

SenseTime’s SenseNova team released SenseNova-Loop-9B-Preview on Oct. 30 under the Apache 2.0 license, its model card says. The model takes text and images. It runs a block of 24 layers twice with shared weights, so about 9 billion parameters act like 14.6 billion at inference, the card says.

Also today

Marvell agrees to buy Lumentum in an all-stock deal. Grok’s API will add hidden watermarks to AI text from Nov. 7, as Claude already does. ChatGPT and Codex will share one usage allowance.