Today in DeAI: Cohere puts hosted inference inside encrypted TEEs, Qwen3.8 open weights land on Hugging Face, and Mozilla pegs the open-model lag at 4.4 months.
Cohere adds confidential computing to Model Vault
Cohere's single-tenant Model Vault now supports confidential computing: workloads run in hardware-encrypted memory (Intel TDX, AMD SEV-SNP, Nvidia confidential computing) with signed attestation tokens, in limited-customer beta since September 16, per VentureBeat. Cohere says neither cloud providers nor itself can read the workloads, and plans to open-source the serving stack for audit — both claims, not yet verified. Why it matters: a checkable middle path between self-hosting and trust-me APIs for sensitive workloads. (VentureBeat) — our coverage
Qwen3.8-27B and Max open weights reach Hugging Face
Qwen3.8-27B weights (BF16 plus blockwise FP8) and the first-ever Max-class open release, Qwen3.8-Max at 2.4T parameters with 95B active, are now on Hugging Face, fulfilling the Qwen team's open-weights promise. The 27B is a dense, vision-capable model with thinking mode on by default; the card notes a hosted QwenCloud version with 1M context is still coming soon. Why it matters: a self-hostable 27B plus a flagship-class open MoE resets the build-vs-rent math for Qwen shops. (Hugging Face, Qwen blog) — our guide
Mozilla: open models trail frontier by 4.4 months
Mozilla's State of Open Source AI report, published September 15, finds Chinese open-weights models lag US closed frontier models by about 4.4 months, and argues most organizations should default to open models — closed buys roughly four months at five times the per-task cost on long tasks, via Ars Technica. Why it matters: the strongest independent number yet for defaulting to open weights and renting the frontier only for the head start. (Ars Technica)
Arcee AI raises $1B-plus Series B for American open weights
Arcee AI announced a Series B led by Vista Equity Partners, Cambium Capital, and Emergence Capital at a valuation above $1 billion, saying it trained four open-weight models for about $20 million and will fund next-generation Trinity models plus DOE lab work. Why it matters: runway for a US open-weights supplier — but builders should watch Hugging Face releases, not the valuation; all figures are company claims. (Arcee AI)
Watching tomorrow
Whether Cohere's open-sourcing pledge gets a date, and whether QwenCloud's hosted Qwen3.8-27B with 1M context opens before third-party hosts set the price conversation.
Sources
- Cohere's Model Vault now encrypts AI inference so even Cohere cannot see enterprise customers' data — VentureBeat
- Qwen/Qwen3.8-27B — Hugging Face
- Qwen3.8-Max: A New Bar for Coding and Cowork — Qwen Team
- Exclusive: Paying for frontier AI models buys 4-month head start — Ars Technica
- Arcee AI Raises Series B to Build the Future of American Open Models — Arcee AI
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.