0G Labs (Home of Infinite AI) (@0G_labs): The best open-source AI model just went private

The best open-source AI model just went private You can now run the #1 ranked open-source model on decentralized infrastructure where nobody logs your prompts. GLM-5, a 744 billion parameter model released under MIT license, ranks #6 overall on the Artificial Analysis Intelligence Index (49.6) and #1 among all open-source models. The five above it are proprietary: Gemini 3.1 Pro, Claude Opus 4.6, Claude Sonnet 4.6, Codex 5.3, and Claude Opus 4.5. It scores 98.2% on agentic tool use (highest of any model, open or closed), and leads the industry in hallucination resistance. And it’s live on 0G Compute’s mainnet at 1 0G (~$0.60) input / 3.2 0G (~$1.92) output per million tokens. Gemini 3.1 Pro charges $2.00 / $12.00. Claude Opus 4.6 charges $5.00 / $25.00. The problem nobody solved until now Open-source AI models have been catching up to proprietary ones for a while. The weights are free. The performance is real. But running them privately? That’s been the gap. Download GLM-5 from HuggingFace, and you still need GPUs. Rent from AWS or Azure, and your prompts flow through someone else’s servers. For teams working with sensitive code or regulated data, that’s been a non-starter. What changed 0G Compute runs GLM-5 inside Trusted Execution Environments. TEEs are hardware-isolated enclaves where even the machine operator can’t see what’s being processed. Your prompts go in, results come out, and nobody in between can access the raw data. Providers on the network do not retain user data after processing. Payments settle through smart contract escrow. Three verification methods (TEEML, OPML, ZKML) provide cryptographic proof that the computation ran correctly. Six models are live on mainnet today. GLM-5 is the flagship, priced at 1 0G (~$0.60) per million input tokens and 3.2 0G (~$1.92) per million output tokens. The benchmarks tell the real story GLM-5 isn’t just “good for open-source.” On the Intelligence Index, it ranks just below Claude Opus 4.6 and Gemini 3.1 Pro. On tau2-Bench (agentic tool use), it scores the highest of any model, open or closed. Codex 5.3 leads Terminal-Bench at $14/M output tokens. GLM-5 also holds the best hallucination score in the industry on the AA-Omniscience Index. On 0G Compute, GLM-5 output costs a fraction of proprietary models. Full benchmark and pricing breakdown in the blog post, but the short version: 84-92% cheaper for a model that leads on agentic tasks. Getting started takes minutes Connect your wallet at compute-marketplace.0g.ai, fund your account with 5 0G, and generate an API key. Then use it with any OpenAI-compatible tool: Works with the OpenAI Python/JS SDKs, Claude Code proxy, MCP server, and 14 agent skills from the 0G Agent Skills repo. What this means for builders GLM-5 on 0G Compute removes the oldest trade-off in enterprise AI: quality vs privacy. The best open-source model, verified computation, zero data logging, on infrastructure no single entity controls. What would you build if your AI couldn’t leak your data? Full analysis from @0G_labs: https://0g.ai/blog/glm-5-live-on-0g-compute