Today in DeAI: Hugging Face handed language-model routing to OpenRouter, Qwen4-Max reached general availability as a closed flagship with open-source siblings promised, and Google Threat Intelligence reported the first observed fully autonomous AI exploit agent.
Hugging Face routes language-model calls through OpenRouter
Every language-model request on Hugging Face Inference Providers now goes through OpenRouter's router, which picks a provider per request on live price, speed and reliability and retries on failure inside the same call. The catch: custom provider API keys stop working for language models, though image, video and audio calls are unchanged. Why it matters: one Hugging Face token now reaches every provider OpenRouter routes to, and the router arbitrages away raw price differences, leaving retention terms and regional hosting as the sticky differentiators. Full story: our coverage (Hugging Face blog).
Qwen4-Max hits GA: closed flagship, open siblings promised
Alibaba's Qwen team made Qwen4-Max generally available: a 2.4-trillion-parameter mixture-of-experts model with 71B active parameters, native vision, a 1M-token context, and a listed price of $1.50/$4.50 per million input/output tokens. Alibaba claims a 62% cost-per-task reduction on its own Work360 workload (a vendor figure, not an independent benchmark), and the launch blog states that smaller models in the family are coming to open source. Why it matters: nothing is open-weight today; the story for self-hosters is the committed follow-up, which given Qwen's track record with permissive 27B-class releases is the line to watch. (Qwen blog)
Google Threat Intelligence reports a fully autonomous AI exploit agent
Google's Threat Intelligence Group says an AI framework found on an exposed server, dubbed exploit_agent, diffed patched versus unpatched code, built exploits and ran them with no human involvement, the first observed case of a fully autonomous exploitation loop. Why it matters: autonomous exploitation collapses the window between disclosure and attack, and it lands hardest on self-hosters and small inference providers, who patch slowest. The finding rests on Google's own reporting; independent replication details are limited to what Google publishes. (The Decoder)
OpenAI starts embedded third-party assessments
OpenAI announced the first phase of embedded assessments: external evaluators Apollo Research, METR, Transluce and Goodfire get embedded access to independently test alignment, safeguards and monitoring on frontier models before deployment. Why it matters: this is a mechanism announcement rather than published results, and no evaluation findings exist yet, but it signals pre-deployment third-party scrutiny becoming part of how a major closed provider ships, relevant context when weighing a closed API against self-hosted weights where you own the safety story. (OpenAI)
Vercel AI Gateway adds Private Inference
Vercel's changelog documents Private Inference: register your own fine-tuned or self-hosted endpoints as private/<slug> providers callable through the same auth, SDKs and logs as catalog models, with optional per-provider Zero Data Retention and CLI provisioning from v65.16.0. Why it matters: it is a managed-gateway on-ramp for open-weight self-hosting, with your endpoint for the sensitive workload, catalog models for the rest, and one audit surface across both. The ZDR toggle is a configuration claim until observed in traffic logs, and endpoint uptime stays the operator's problem. Full story: our coverage (Vercel changelog).
Watching tomorrow
Akash's compute-marketplace cutover vote is the next governance event in the network's two-year Solana migration, and running workloads keep Cosmos continuity for 90 days after any cutover, which is worth tracking if you serve inference on Akash.
Sources
- Hugging Face and OpenRouter join forces to make open-weight models easy to run — Hugging Face
- Qwen4-Max GA announcement — Qwen Team (Alibaba)
- Google warns that fully automated cyberattacks have arrived and will get worse fast — The Decoder
- Our first phase of embedded assessments — OpenAI
- AI Gateway now supports Private Inference — Vercel
- Akash Network's AKT goes live on Solana as Cosmos migration begins — crypto.news
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.
