A few days ago, Hugging Face suffered an autonomous AI attack-the attacker used OpenAI's agent to penetrate the platform, methods隐蔽. The community wanted help with security response, but US closed-source commercial models, due to built-in safety rails, directly refused to assist in analyzing malicious code and attack traces.
The one that finally completed the forensic analysis was China's open-source model GLM. The dev team locally deployed GLM on their own infrastructure-no data leak risk, no ethical constraints blocking them-directly processed the malicious code and reconstructed the attack chain.
This made many reconsider the claim that "closed-source equals security."
Open Source Holds 62.8% by Count-Performance and Cost Both Reverse
Per Introl's late-2025 data, open-source models hold 62.8% market share by count-two years ago proprietary models dominated. Alibaba's Qwen3-235B-A22B matches or exceeds GPT-4o on most public benchmarks, activating only 22B of its 235B params per run-highly efficient. Meta's Llama 4 Scout packs a 10M-token context window into a single H100 GPU, runs on INT4 quantization, swallowing 7,500 pages at once.
Cost has flipped entirely. DeepSeek V3.2 charges $0.26 per million input tokens-about a tenth of GPT-5.2 Pro. Mistral Medium 3 reaches 90% of Claude Sonnet 3.7's performance at 8× lower price. Organizations report mixed open-source strategies yield 25% higher ROI than pure closed-source. This isn't "cheap means bad"-it's "expensive isn't necessarily justified."
Medical institutions that spent heavily fine-tuning专用 models on old base models got crushed by continuously-updated open-source general LLMs in real clinical tests. Non-English medical records garbled, multimodal dirty-data fusion a mess. Wrong route, no amount of effort helps.
Why Open Source Reverses: Engineering Law Reconstruction
On iteration speed, open-source collective wisdom outpaces any single closed-source company's internal R&D. 2026 Lunar New Year: Kimi K2.5, Qwen3.5, GLM-5, MiniMax M2.5 collectively erupted-Agent Swarm, multimodal, 60% cost drops-things no single company can do all at once.
Data sovereignty and deployment flexibility are key. When your model must run on someone else's API, you're always in their rules. Self-hosted open-source lets hospitals run patient-data queries locally, financial firms fully control trading algorithms-no worry about vendors suddenly deprecating old versions, no usage-based pricing hijack.
And the guardrail paradox. Closed-source models pile on guardrails for "safety," but in real security-response scenarios these guardrails become obstacles. GLM could complete forensics precisely because it lacks those "politically correct" ethical constraints. In some scenarios, open-source is the only viable choice.
For Regular People: One Machine Is a Software Company
For regular people, you no longer need to beg giants for API quota to use cutting-edge AI.
Deploying a 10M-token-context model on your own computer costs negligibly. The minimind project lets you run the full pretrain-to-SFT pipeline for 3 yuan; open-source AI coding tools like OpenCode have 6.5M+ monthly active users. One person, one machine, roughly a software company.
More importantly, the future belongs to those who can master open-source models-not "using AI," but building your own AI workflows. The open-source model that warns you in the ER at midnight may be the one you personally tuned during the day.
References
- Unprecedented autonomous AI attack, US closed-source models refuse to assist - Guancha
- Hit by OpenAI model "autonomous attack," world's largest AI open-source community completes forensics with Chinese AI - Sina
- Open-source AI models close the gap: DeepSeek, Qwen3, and Llama 4 now rival GPT-5 - Introl
- 2026 Lunar New Year, Chinese open-source LLMs collectively erupt - AI Coding
- 2026 Personal Guide to Training AI LLMs from Scratch - Unifuncs