LongCat-2.5 explained: native multimodality, 1M context and agent gains over 2.0

What matters in this release?

Meituan’s LongCat team released LongCat-2.5 on November 1, 2026. Its main change from LongCat-2.0 is the combination of native visual understanding with longer, multi-step agent workflows: reading interfaces and charts, calling tools, editing code and checking the result.

This article draws on the official announcement and API documentation, checked on November 2, 2026. All benchmark results below are publisher-reported; TboxIO has not independently reproduced them.

Specifications and access

Item	LongCat-2.5
Total / active parameters per token	1.6 trillion / about 48 billion
Main capabilities	Native visual understanding, coding, long-horizon agents and tool use
API context limit	1,048,576 tokens, including input and output
API max_tokens limit	262,144 tokens (256K)
API compatibility	OpenAI Chat Completions, Responses and Anthropic Messages formats
Weight availability	GitHub and Hugging Face links marked Coming soon on the release page

A 1M context window is a total per-request capacity, not a promise of 1M generated tokens or unchanged accuracy at full capacity. The 256K output parameter ceiling is also constrained by remaining context and request settings.

Three changes from 2.0

Vision uses a shared token backbone. Residual vector quantization (RVQ) converts visual features into multiple levels of discrete tokens, allowing visual and text signals to share a backbone. Target workloads include screenshot-based web development, chart analysis and 3D coding. Understanding images and writing visual code do not establish that the API can natively generate images or video.

Training emphasizes complete task trajectories. Longer-sequence training and long-horizon reinforcement learning incorporate reproducible tasks derived from real coding failures, repository issues, test-based validation and multiple agent harnesses. The intended benefit is more reliable planning, execution, inspection and repair across a workflow.

Serving infrastructure adds W4A4. With both expert weights and activations at 4 bits, the team reports 43.3% higher prefill throughput, 42.77% higher decoding throughput and 15.91% lower time to first token against a W8A8 baseline. These are infrastructure stress-test results under particular conditions, not a universal API speedup over LongCat-2.0.

Reported results versus LongCat-2.0

Benchmark	LongCat-2.0	LongCat-2.5	Score change
Terminal Bench v2.1	70.8	88.4	+17.6
Terminal Bench v4.0	3.0	38.5	+35.5
DeepSWE v1.1	9.7	64.6	+54.9
PaperBench	45.2	82.1	+36.9
BrowseComp	79.9	91.9	+12.0
AutomationBench v1.0.6	15.8	51.8	+36.0
MCPAtlas	61.8	83.1	+21.3
JobBench	44.4	65.4	+21.0

Changes are absolute score differences, not relative percentage increases. Scores from different benchmarks should not be added or directly compared. Terminal Bench v2.1 and v4.0 are different tests.

The reported gains are substantial in complex code execution, research reproduction, browsing and multi-tool work. Harness details matter: Terminal Bench uses Claude Code v2.1.280 with a 512K sequence limit; DeepSWE uses mini-swe-agent with a 1M limit; BrowseComp uses a custom harness with context-clearing strategies. These results measure a model-and-tool configuration together.

How does it compare with other models?

The release table lists Terminal Bench v2.1 scores of 88.4 for LongCat-2.5, 88.8 for GPT-5.6-Sol, 89.1 for Claude Opus 5 and 90.6 for DeepSeek-V4.1-Flash. On BrowseComp, it lists 91.9 for LongCat-2.5, 91.2 for Kimi-k3 and 91.5 for DeepSeek-V4.1-Flash.

Those figures suggest competitiveness on selected long-horizon tasks, not an overall lead. Some competing results come from their own public reports, while others were run by the LongCat team. This is not an independent comparison under one uniform setup. Re-run your own tasks with the same repositories, prompts, tool permissions and budget before switching models.

For vision, the team reports 78.6 on Vision2Web, 90.2 on CharXiv-RQ with tools and 85.2 on VideoMME. These cover different web, chart and video-understanding tasks; their scores are not interchangeable or direct measures of business accuracy.

API pricing and weight availability

At the time of checking, the official pricing page explicitly names LongCat-2.5 rather than only Preview. Limited-time pay-as-you-go rates, in USD per million tokens, are:

Billing item	Limited-time rate
Uncached input	$0.30
Cached input	$0.006
Output	$1.20

Rates may change; the platform and actual billing records take precedence. Long agent runs can accumulate repeated input, output and tool calls, so the cost of one prompt can underestimate the total task cost.

Online access and API documentation are available, but the announcement still marks its weight repositories Coming soon. A released model should therefore not be described as already having downloadable weights, a confirmed license or an established local deployment requirement.

TboxIO’s assessment

The most useful initial trials are screenshot-driven development, edits across multiple files, and research or office workflows with repeated tool use. The combination of stronger reported long-horizon performance and low promotional token rates makes it worth evaluating.

Use tasks with explicit acceptance criteria and record completion rate, manual rework, end-to-end latency and total cost per task alongside your current model. Keep tests and human review for sensitive charts, financial work and code changes: a large context window and high benchmark scores do not guarantee correctness.