Frontline Hotspot
Frontline Hotspot

AI Open-Source Reversal: 62.8% Market Share, 1/10 Cost, Hugging Face Attack Forensics via GLM

Hugging Face hit by autonomous AI attack; US closed-source models' safety rails refused to assist; China's open-source GLM locally deployed completed forensics. Open source holds 62.8% market, Qwen3/Llama 4 rival GPT-4o, DeepSeek at 1/10 cost. Guardrail paradox + data sovereignty + iteration speed drive the reversal.

Published July 26, 20264 min read
<!-- ai-open-source-reversal-hotspot | hotspot | AI Open-Source Reversal -->

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

This article is AI-assisted and human-edited. Last updated: 2026-07-26

FAQ

Are open-source AI models now stronger than closed-source?
Performance is close (Qwen3-235B rivals GPT-4o, Llama 4 Scout 10M tokens), cost a tenth (DeepSeek V3.2 vs GPT-5.2 Pro), market share 62.8%. In vertical scenarios like healthcare, open-source general LLMs crush specialized closed-source. But closed-source still leads in top creative/complex reasoning.
Why was GLM used for Hugging Face attack forensics?
US closed-source models' safety rails refused to analyze malicious code/attack traces. GLM is open-source, locally deployed with no data leak + no ethical constraints blocking, directly processing malicious code to reconstruct the attack chain. Guardrail paradox: closed-source "safety" becomes an obstacle in real security scenarios.
How can regular people use open-source models?
Local deployment (minimind runs full pretrain-to-SFT for 3 yuan, OpenCode has 6.5M MAU). Self-hosted data stays on-domain, no API usage-based pricing hijack. The future belongs to those who master open-source models to build their own AI workflows.

Related