美团以MIT许可开放LongCat-2.5多模态大模型权重

Summary 1: Meituan released LongCat-2.5 weights under the MIT license, allowing developers to use them commercially, modify them, and redistribute them; the release came just one week after the model's initial launch.

1. Capabilities and scenarios: The model has visual understanding and can convert screenshots, charts, images, and videos into tokens for joint processing with text. It is suited to codebase analysis and multi-step software writing and revision; a typical use case is using a screenshot of a broken checkout flow to locate the relevant repository components, generate a fix patch, and verify it in a browser.

2. Interface and examples: The API can embed images in requests or pass images and videos via URL; the video interface supports URL only. Existing front-end examples include an auto-launch website, a fantasy game interface, and a store-management site.

3. Data and cost: Terminal Bench v2.1 score 88.4 versus 70.8 for the prior generation; DeepSWE v1.1 score 64.6 versus 9.7; SWE Atlas score 58.3 versus 33.1. Non-cached input costs $0.30 per million tokens, output $1.20, and cached input $0.006; a single run with 200,000 input tokens and 100,000 output tokens costs $0.18.

4. Deployment caveats: Production pages still need functional, accessibility, and maintainability checks. Testing should standardize repository state, browser environment, test suite, tool permissions, and time limits; screenshot-to-code should separately evaluate rendering and code quality; bug fixes should verify root cause and unrelated functionality.

Summary 1: The brand-and-ecosystem strategy is clear. Meituan released the weights under the MIT license and uploaded the Hugging Face checkpoint one week after launch, allowing commercial use, modification, and redistribution. This quickly attracts developers and strengthens its technology brand and ecosystem influence.

1. Product development: LongCat-2.5 has visual understanding and supports screenshots, charts, images, and videos as inputs. It targets codebase analysis and multi-step software writing and revision, creating a product selling point that combines visual judgment with coding capability.

2. Performance endorsement: Terminal Bench v2.1 score 88.4 versus 70.8 for the prior generation; DeepSWE v1.1 score 64.6 versus 9.7; SWE Atlas score 58.3 versus 33.1. It is within 1 point of GPT-5.6 Sol and Claude Opus 5 on Terminal Bench and leads its peer tier on 3DCodeBench.

3. Pricing competition: Non-cached input is $0.30 per million tokens, output $1.20, and cached input $0.006; a single run costs $0.18. Compared with DeepSeek at $0.09, GPT-6.1 Sol at $1.40, Claude Opus 5.5 at $2.80, Claude Fable 5.2 at $5.60, and GPT-6 Astra at $7.00, the price advantage can be used as a brand messaging point.

4. Marketing caveats: The closed-source versions used for comparison predate the latest market versions, and differences in test settings prevent fully like-for-like comparison. Front-end generation examples can be used for demonstration, but production deployment requires functional, accessibility, and maintainability checks.

Summary 1: The opportunity lies in the MIT license, which permits commercial use, modification, and redistribution, plus limited-time hosted API discounts. Non-cached input is $0.30 per million tokens, output $1.20, cached input $0.006; a single run with 200,000 input and 100,000 output tokens costs $0.18. This is lower than most competitors but higher than DeepSeek's off-peak $0.09.

1. Demand changes: Multimodal vision plus coding can locate components from failure screenshots, generate patches, and verify them in a browser. Front-end generation can produce an auto-launch website, a fantasy game interface, and a store-management site, creating opportunities for internal tools and business pages.

2. Risk warnings: The video interface is URL-only, limiting internal screen-recording teams; production pages require functional, accessibility, and maintainability checks; private code screenshots require confirmation of the hosted service's data retention, access, and deletion rules, and should first be validated with public or synthetic data.

3. Policy interpretation: The United States is scrutinizing Chinese AI models; DHS requires federal civilian agencies to exclude DeepSeek from procurement and deployment, and the Department of Defense's supply-chain risk determination on Anthropic takes effect within days. No provision currently names Meituan or LongCat-2.5, but serving U.S. government customers requires verifying contracts and the latest federal supply-chain directives, preparing fallback models, and noting that private deployment or U.S.-based cloud may also be restricted.

4. Procurement accounting: Enterprises should calculate the total cost per accepted result, factoring in retries, sandbox duration, fallback models, and engineer review time. Cache hit rate and token consumption affect spending.

Summary 1: The digitalization opportunity is clear. Visual understanding combined with coding capability can support production-system maintenance: using a screenshot of a broken checkout flow to locate repository components, generate a fix patch, and verify it in a browser, while also supporting codebase analysis and multi-step software writing and revision.

1. Design and production needs: It supports screenshots, charts, images, and video inputs and can convert visual information into tokens for joint processing with text. It is suitable for generating front-end content such as factory internal tools, equipment interfaces, and store-management sites.

2. Deployment requirements: The MoE model has 1.6 trillion to 1.8 trillion total parameters, activates about 48 billion per token, and has a 16-bit weight file of about 3.2TB. Meituan published an SGLang-FluentLLM deployment guide for neural processing units; the document says domestic accelerator cards have achieved 4-bit compression of expert weights and activations.

3. Implementation path: For occasional prototyping, hosted API should be preferred. Teams with heavy visual-coding workloads and high data sensitivity should then evaluate private deployment and verify loading compatibility, visual understanding and tool-calling capability, concurrency latency, and load fluctuation.

4. Business implications: The MIT license permits commercial use, modification, and redistribution; the API costs $0.18 per run, enabling low-cost demos and internal tools. Production pages still require functional, accessibility, and maintainability checks.

Summary 1: From an industry-trend perspective, Meituan released LongCat-2.5 weights under the MIT license and opened the checkpoint one week after launch. Multimodal visual-plus-coding models have entered a stage where they can be used commercially and redistributed, with hosted APIs and private deployment running in parallel.

1. New technology: Visual information is converted into tokens and processed jointly with text, supporting screenshots, charts, images, and videos; video is URL-only. The MoE model has 1.6 trillion to 1.8 trillion total parameters and activates about 48 billion; the SGLang-FluentLLM deployment guide notes that domestic accelerator cards use 4-bit compression for expert weights and activations.

2. Customer pain points: The URL-only video interface limits internal screen-recording teams; production code requires functional, accessibility, and maintainability checks; private code screenshots require confirmation of data retention, access, and deletion rules; private deployment requires about 3.2TB of weights, and other compute clusters must independently verify compatibility and performance.

3. Solutions: Use the hosted API for occasional prototyping; evaluate private deployment only when visual-coding workloads are heavy and data sensitivity is high. For testing, standardize repository state, browser environment, test suite, tool permissions, and time limits, and run repeated tests to measure success rates.

4. Pricing and compliance: Non-cached input is $0.30 per million tokens, output $1.20, and cached input $0.006; a single run costs $0.18. Enterprise procurement should account for retries, sandboxing, fallback models, and review costs. Providers serving U.S. government customers should verify federal supply-chain directives and prepare alternative models.

Summary 1: The platform's latest move is that Meituan uploaded the LongCat-2.5 checkpoint to Hugging Face just one week after launch, releasing the weights under the MIT license and permitting commercial use, modification, and redistribution. It also offers limited-time hosted API discounts, creating a commercial combination of open weights plus API.

1. Developer demand: The model has visual understanding and supports screenshots, charts, images, and video inputs. It targets codebase analysis and multi-step software writing and revision. Existing front-end examples include an auto-launch website, a fantasy game interface, and a store-management site, which can serve as platform recruitment and demo materials.

2. Operations management: The API can embed images or pass images and videos via URL; video is URL-only. Production pages require functional, accessibility, and maintainability checks; private code screenshots require confirmation of data retention rules; private deployment requires about 3.2TB of weights, and other compute clusters must verify compatibility and concurrency latency.

3. Regulatory risk avoidance: The United States is scrutinizing Chinese AI models; DHS requires federal civilian agencies to exclude DeepSeek from procurement and deployment, and the Department of Defense's supply-chain risk determination on Anthropic takes effect within days. No provision currently names Meituan or LongCat-2.5, but suppliers to government customers must verify contracts and the latest federal supply-chain directives, prepare fallback models, and note that private deployment or U.S.-based cloud may also be restricted.

4. Pricing strategy: Non-cached input is $0.30 per million tokens, output $1.20, and cached input $0.006; a single run costs $0.18. This offers a price advantage versus DeepSeek at $0.09, GPT-6.1 Sol at $1.40, Claude Opus 5.5 at $2.80, among others.

Summary 1: A new industry development is that a Chinese AI model has released multimodal large-model weights under the MIT license. Meituan uploaded the LongCat-2.5 checkpoint one week after launch, permitting commercial use, modification, and redistribution, with open weights and a hosted API in parallel.

1. Technical architecture: Visual information is converted into tokens and processed jointly with text, supporting screenshots, charts, images, and videos; video is URL-only. The MoE model has 1.6 trillion to 1.8 trillion total parameters, activates about 48 billion per token, and has 16-bit weights of about 3.2TB. The SGLang-FluentLLM deployment guide notes that domestic accelerator cards use 4-bit compression for expert weights and activations.

2. Performance and evaluation issues: Terminal Bench v2.1 score 88.4 versus 70.8 for the prior generation; DeepSWE v1.1 score 64.6 versus 9.7; SWE Atlas score 58.3 versus 33.1. It is within 1 point of GPT-5.6 Sol and Claude Opus 5 on Terminal Bench and leads on 3DCodeBench. However, DeepSeek-V4.1-Flash scores highest on DeepSWE, Claude Opus 5 is best on Vision2Web, and Claude scores 53.9 on Terminal Bench v4.0 versus LongCat's 38.5. The comparison versions predate the latest releases, and differences in test settings prevent fully like-for-like comparison.

3. Policy implications: The United States is scrutinizing Chinese AI models; DHS excludes DeepSeek from federal civilian procurement and deployment, and the Department of Defense's supply-chain risk determination on Anthropic takes effect within days. No current provision names Meituan, but third-party deployments may also be restricted.

4. Business model: The MIT license plus limited-time hosted API pricing: non-cached input is $0.30 per million tokens, output $1.20, and cached input $0.006; a single run costs $0.18. Compared with DeepSeek at $0.09, GPT-6.1 Sol at $1.40, and Claude Opus 5.5 at $2.80, enterprise procurement must calculate total cost.