Introducing LongCat 2

Pippa A model that can build a fluid simulation, a three-dimensional exhibition, and a whole software project is aiming at the part of AI work that usually breaks first: staying useful after the impressive opening move.

Tyler That’s the pitch for LongCat-2.5. It’s a native multimodal model for long-horizon agents, with 1.6 trillion total parameters and roughly 48 billion active for each token. The interesting bit is less the headline size than how much of the training system is built around unfinished work.

Pippa Right.

Tyler The team says it trained on longer agent trajectories, then used large-scale reinforcement learning across complex codebases and full-project builds. In plain English, the model gets feedback from environments while it works, rather than only learning from completed answers.

Pippa And that gives it a much clearer user story than another model that just writes a nicer paragraph. The target is somebody handing over a visual interface, a repository, or a multi-step engineering task and wanting the agent to keep its footing through the boring middle.

Tyler Exactly. LongCat-2.5 also brings vision into the same token-based backbone as text. Its visual system turns images into multiple levels of discrete tokens, which is a little like giving pictures their own vocabulary before mixing them with words.

Pippa That matters for the demos. A model can look at a design or a visual reference, reason about it, and generate code for an interactive page without vision being a separate bolt-on service. The launch specifically points to real-time fluid simulation, visual effects, three-dimensional scenes, and office workflows.

Tyler The implementation has three practical pieces. A native-resolution vision transformer preserves detail at different image sizes. A fixed tokenizer can preprocess and cache visual tokens offline. Then the visual embeddings are learned jointly with the language model, so the two modalities can occupy a shared semantic space.

Pippa Mm-hm.

Tyler That could simplify serving and training, especially for long sequences. But the source is careful in one place: the shared-space evidence comes from a smaller mixture-of-experts model with about three billion active parameters. That’s useful validation, not proof that the full system behaves perfectly.

Pippa I appreciate that distinction, because the product demos are doing a lot of emotional work here. They make the model feel like a designer, engineer, and analyst in one sitting. The adoption question is whether those capabilities survive a real project with unclear requirements, changing files, and tests that catch the model cheating.

Tyler And LongCat’s training pipeline is unusually explicit about that problem. It filters pull requests and linked issues, synthesizes environments, and checks tasks with fail-to-pass and pass-to-pass tests. It also turns real agent failures into reproducible training cases instead of treating every bad trajectory as an anecdote.

Pippa That’s the part I find genuinely promising. A failed coding attempt becomes a testable object the team can train against, which creates a path from user pain to model improvement. It’s a much more concrete upgrade loop than saying the agent has better reasoning.

Tyler Sure. There are defenses against the obvious gaming too. The environments lose network access, git history, and cached files. An LLM judge audits trajectories for exploits, while penalties target repetitive reasoning and repeated tool calls. They also adjust rewards to discourage both endless overthinking and unnaturally short answers.

Pippa And they mix tasks and scaffolds during reinforcement learning, so the model isn’t trained for one harness and then celebrated for surviving only that harness. That connects to the long-horizon pattern we’ve been watching since Inkling 2 Small, where the harness robustness was more interesting than the parameter count.

Tyler Yeah.

Pippa Zooming out, the launch compares LongCat-2.5 with GPT-5.5, GPT-5.6 Sol, Claude Opus 4.8, and Claude Opus 5. But the material here gives us the comparison set, not enough detailed results to crown a winner, so I’d treat the dashboard as an invitation to test rather than a verdict.

Tyler Same. The clever engineering is in the whole loop: mixed scaffolds, asynchronous rollouts, controls for stale policy updates, and infrastructure that keeps long-context training within less than 80 gigabytes per domestic accelerator. The questionable part is simply what we can’t yet verify from this launch page: how much of that translates to independent, messy workloads.

Pippa For trying it, the source points to LongCat’s chat experience and API access. GitHub and Hugging Face are marked as coming soon in the article, so the upgrade path today looks more like testing the hosted surface than immediately self-hosting the full model.

Tyler And that makes the next useful comparison pretty concrete: give it the same repository, visual brief, test suite, and time budget as the other agent models, then inspect where it fails. Not just whether the final demo looks good, but whether it preserves requirements and avoids burning the whole context on repeated tool calls.

Pippa That’s a very respectable amount of machinery for one launch. Tyler, I’m going to keep the fluid simulation tab open and pretend that counts as research. See you next time.