This week's local AI releases deserve a procurement test

A model license, a laptop performance claim and an inference-library release answer different questions. Treating them as one breakthrough is tempting. Treating them as separate checks is more useful for a company deciding whether to run AI on hardware it controls.

This week supplied an example of each. Meituan announced an MIT-licensed multimodal model. Edge0 reported a large model running on a compatible laptop. The llama.cpp maintainers released version 1.0.0 and explained how they will signal interface changes. Together, these are reasons to investigate local deployment, not proof that it will beat a hosted service for a particular business.

Start with permission. Meituan's November 10 announcement says LongCat-2.5 is available under the MIT license for commercial inference deployment and further training without extra authorization. Its published license permits use, modification and redistribution, subject to retaining the copyright and permission notice. That is a concrete input to a procurement review. It is not a service-level agreement: the same license provides the software "as is," without warranty.

Meituan describes LongCat-2.5 as a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion activated per token, designed for long-running multimodal tasks. Those two parameter counts should not be confused. Activating fewer parameters for each token does not make the rest of the model disappear from the deployment plan. A buyer still needs a supported serving setup and a measured hardware budget. Nor should access to weights be confused with access to every ingredient of training.

Next comes fit. In its November 12 release, Edge0 reports that Edge0-145B reaches more than 80 decoding tokens per second with 29GB peak memory on a compatible laptop. These are company measurements, not an independent guarantee. Its comparison discusses a 48GB laptop, and the post also describes system-memory and SSD tiers. A headline memory number alone is therefore not a complete shopping specification. The company says the full hardware environment and reproduction instructions are being released with its technical report. Procurement should ask for those details and repeat the test on the intended machine, with realistic input lengths and concurrent applications.

Finally, consider maintenance. The llama.cpp team's November 13 post explains why version 1.0.0 matters to downstream developers: major-version bumps will signal public interface incompatibilities. The release notes also list server hardening and memory-management changes. That is useful engineering information, not a certification that an application is secure. A team embedding the library still needs an update owner, regression tests and a plan for breaking changes.

My recommendation is to turn these announcements into a small, bounded comparison. Choose one internal workload with clear acceptance criteria. Run the local candidate and the hosted alternative on the same permitted inputs. Measure output quality, response time, failures and the work needed to operate each option. Include hardware, electricity, maintenance and support in the cost estimate, rather than comparing a download price with an API invoice.

Give data handling its own test. Local inference can keep model inputs on controlled hardware, but an application may still call search, telemetry, storage or other online services. Inspect those connections before promising that nothing leaves the building. License freedom and network isolation are separate properties.

The purchasing question is not whether open models have won. It is whether a particular deployment meets the business's quality, cost and control requirements. This week's releases offer new candidates and useful infrastructure commitments. The next evidence should come from the buyer's own test, not a larger parameter count or a more confident launch headline.

Sources

• Meituan, LongCat-2.5 open-source announcement (November 10, 2026): https://tech.meituan.com/2026/11/10/Meituan-LongCat-2.5-Open-Source.html
• Meituan, LongCat-2.5 model card: https://huggingface.co/meituan-longcat/LongCat-2.5
• Meituan, MIT license text: https://huggingface.co/meituan-longcat/LongCat-2.5/raw/main/LICENSE
• Edge0, Edge0-145B release (November 12, 2026): https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer
• ggml-org, llama.cpp v1.0.0 announcement (November 13, 2026): https://huggingface.co/blog/ggml-org/llama-cpp-v1
• ggml-org, llama.cpp v1.0.0 release notes: https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0

Full disclosure: Ganesh uses Instinct (instinct.ai), an AI assistant, to research, write and publish this blog series.