Today in DeAI: Microsoft ships MIT-licensed open weights for a 2.3B model that is now Edge's built-in, Cloudflare opens the Clairic extraction weights, and a NEAR AI incident report details what its TEE didn't protect.
Microsoft takes Aion 1.0 Instruct GA with MIT open weights and an Edge Prompt API trial
Microsoft published open weights for Aion 1.0 Instruct, a 2.3B-parameter model, under the MIT license on Hugging Face, and made it the built-in small language model of Microsoft Edge, replacing Phi-4-mini. The Prompt API moved to an origin trial so live sites can call on-device inference, and an experimental Decisions API entered preview, answering classification questions in a single pass without text generation. Microsoft's self-run benchmarks put MMLU at 55.15 and IFEval at 83.90, and no independent reproduction exists yet. Why it matters: an MIT-licensed open model now sits on the default inference path of a mainstream desktop browser, which makes the on-device tier of hybrid local-plus-cloud inference something sites can actually build against. (Microsoft Edge Blog) Our coverage
Cloudflare publishes open weights for Clairic 27B and Clairic-flash 9B extraction models
Cloudflare put full open weights for both extraction models on Hugging Face and hosts them on Workers AI at $0.24 and $0.10 per million input tokens respectively, billed on input tokens only, with both at the promotional $0.10 rate until December 1. The models pull dates, numbers, tables, and confidence spans out of text, images, and PDFs, and the company's latency and accuracy tables are explicitly self-reported. Why it matters: documents teams hesitate to send to third-party APIs now have a downloadable local path, and the input-only billing sets a concrete price point for extraction-grade work. (Cloudflare Blog) Our coverage
NEAR AI report: SGLang flaw exposed ~100 users, and the 18 affected agents weren't in the TEE
NEAR AI's incident report says an SGLang remote-code-execution flaw let an attacker expose about 100 Private Chat users' account details, chat history for 18 legacy agents, and two API keys since revoked. The report also concedes the enclave doesn't protect against a vulnerability in the software running inside it, and the affected agents were not running inside the TEE. All breach details are provider self-reports with no independent forensics. Why it matters: enclave marketing reduces but does not remove provider-side exposure, and the serving stack itself stays in the attack surface either way. (MLQ.ai) Our earlier coverage of the incident report
Perplexity launches Orchestrator API with 45+ models behind one endpoint
Perplexity's Orchestrator routes more than 45 open-weight and proprietary models through one endpoint with same-model provider failover, unified billing, and an automatic router, stepping into the territory OpenRouter and Vercel AI Gateway occupy. The routing and failover claims are self-reported. Why it matters: a second at-scale routing layer means more leverage on price and redundancy for builders, and one more intermediary whose logging and retention terms need reading before anything sensitive goes through it. (Perplexity)
Ramp AI Index: token prices down 21% in six weeks while usage hits records
Ramp's spend-panel index reports token prices falling 21% in six weeks, record usage, and Anthropic taking roughly 59% of measured token spend, with The Decoder corroborating the pattern against Stripe and OpenRouter data. These are third-party spend estimates with methodology caveats, not provider figures. Why it matters: price deflation this fast makes long-term commitments to any single provider's list price a bad bet, which is the portable-workload argument in numbers. (The Decoder)
Watching tomorrow
Baidu's ERNIE 5.5 Flash open weights are rumored for around November 25, so expect the verification pass to start with Baidu's own channels and its Hugging Face org.
Sources
- Hybrid intelligence on the Web: Introducing Aion Instruct and Prompt API in Microsoft Edge — Microsoft Edge Blog
- An updated Clairic and Clairic-flash are now open-weight — Cloudflare Blog
- NEAR AI says an SGLang flaw let an attacker expose about 100 users' account details — MLQ.ai
- Introducing Orchestrator — Perplexity
- AI industry growth remains strong, but the pace is slowing — The Decoder
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.
