本週,「擁有自己的 AI」從口號變成企業採購選項
The week "own your AI" stopped being a slogan
llama.cpp 發佈 1.0、Meituan 公開 LongCat-2.5 權重,Edge0 亦推出 Edge0-145B,文章認為本地 AI 堆疊正成為企業租用雲端模型以外的採購選項。
On Friday, a piece of software most executives have never heard of quietly hit version 1.0. It is called llama.cpp, and it is the open-source engine that lets organizations run large AI models on their own hardware - no API keys, no per-token invoices, no data leaving the building. A 1.0 tag is a ceremony, not a feature, but look at what the maintainers chose to ship with it: server hardening against SSRF, CORS and DNS-rebinding attacks, an OpenAI-compatible speech endpoint, and a cache that lets a laptop offload parts of giant mixture-of-experts models to CPU memory. Those are enterprise concerns, not hobbyist ones. The local-AI stack is dressing for work.
The engine matters because the models arriving for it are no longer toys. On Monday, Meituan - the Chinese delivery and e-commerce giant - published the weights of LongCat-2.5 on Hugging Face under the MIT license, as VentureBeat first reported and the model card confirms. This is a 1.6-trillion-parameter mixture-of-experts model (about 48 billion parameters active per token), built natively for multimodal, multi-step agentic work: reading screenshots, working through codebases, executing long tasks. Meituan's own benchmark table shows it competitive with frontier proprietary models on several coding and agent benchmarks, though trailing the best of them - Claude Opus 5 and DeepSeek V4.1 - on Terminal Bench v2.1. The honest reading is not "open beats closed." It is that a genuinely permissive license now sits near the frontier, and the gap is measured in single-digit points on the lab's own scoreboard.
Then on Wednesday, a startup called Edge0 released Edge0-145B, which it describes as the first open-source model of that size aimed squarely at personal hardware. The company's own numbers claim 145 billion parameters running on a laptop within 29GB of peak memory at over 80 tokens per second, with a smaller sibling running on a phone. Those figures are vendor-reported and await independent testing, and the chart behind them - 67 open models of at least a billion parameters from 14 labs - is also the company's own assembly. But even discounted, the direction is the story: the on-device ceiling, stuck under 10 billion parameters before this September by Edge0's account, is being pushed up by an order of magnitude.
Why should a leadership team care? Because "build vs. buy" in AI just acquired a third answer: own. The default enterprise AI posture - rent a frontier model from a cloud provider, per seat and per token - now has a credible alternative for a widening band of workloads. A permissively licensed model you can inspect, fine-tune and run where your data lives changes the negotiation even if you never deploy it: it is a floor under vendor pricing and a ceiling on lock-in. The CIO's question is shifting from "which vendor do we choose?" to "which workloads actually need a rented frontier model, and which can run on something we control?"
Three cautions belong in the same breath. First, the benchmark numbers here are self-reported; LongCat-2.5's own table is careful to flag which scores are measured in-house. Second, open weights shift responsibility, not just cost - security patches, evaluation and guardrails become your problem, which is exactly why llama.cpp's 1.0 hardening work matters. Third, the frontier is not standing still; the same week proved the proprietary labs still hold the top spots on the hardest agentic tests.
But the strategic picture changed this week. A trillion-parameter-class model under an MIT license on Monday, a 145-billion-parameter laptop model on Wednesday, and the local engine's 1.0 on Friday. Owning your AI stack used to be an ideological position. It is becoming a procurement option - and options are leverage.
Sources
- VentureBeat, "Meituan releases the open weights of its LongCat-2.5 LLM with vision under a permissive MIT License" (November 9, 2026): https://venturebeat.com/technology/meituan-releases-the-open-weights-of-its-longcat-2-5-llm-with-vision-under-a-permissive-mit-license
- Meituan-LongCat, "LongCat-2.5 model card" (Hugging Face): https://huggingface.co/meituan-longcat/LongCat-2.5
- Edge0, "Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer" (November 12, 2026): https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer
- ggml-org, "llama.cpp v1.0.0 release notes" (November 13, 2026): 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.
來源:DEV Community · dev.to