DeAI Daily Brief — 11 September 2026

Today in DeAI: NEAR AI Cloud put verifiable TEE private inference on an OpenAI-compatible API, DeepSeek shipped the MIT-licensed V4.1-Flash and repriced around it, and IBM and NASA open-sourced a lunar foundation model.

NEAR AI Cloud runs open models inside Intel and NVIDIA enclaves

NEAR AI Cloud is serving TEE-hosted open-weight models — GLM, Qwen, and others — inside Intel TDX confidential VMs and NVIDIA H100/H200 GPU enclaves, with TLS terminating inside the enclave and every response signed by a key that never leaves the hardware. Why it matters: the privacy claim moves from policy ("we don't look") to a cryptographic attestation users can check, though the check itself is still on the builder to run — we have not independently verified it. (NEAR AI docs) — our coverage

DeepSeek releases V4.1-Flash under MIT, cuts cached input to $0.003/M

DeepSeek published a 552B-parameter mixture-of-experts with a new causal encoder-decoder architecture, 1M-token context, and multimodal support — weights public under MIT. The API's deepseek-flash already serves it, cached input is $0.003 per million tokens off-peak, and from September 14 all deepseek-v4-pro traffic routes to it at Flash prices. Why it matters: if the efficiency claims hold, serving economics for long-context open models step down again; benchmark claims are still DeepSeek's own. (model repo, pricing) — our PULSE on the discourse

IBM and NASA open-source a lunar foundation model

The NASA-IBM Lunar Foundation Model is public on Hugging Face under Apache 2.0, trained on multi-instrument lunar observation data. IBM says it cuts error in identifying potential lunar ice deposits by up to 22% (RMSE) against a SwinV2-B baseline. Why it matters: it is one of the first openly licensed foundation models built for planetary science — a template for open scientific models outside the chat-model rat race. (IBM Newsroom, model repo)

Abacus.AI ships three open-weight Smaug fine-tunes

Abacus.AI released three Smaug models for agentic workloads, fine-tuned on open-weight bases from the Qwen, DeepSeek, and Kimi families. Why it matters: another data point that open bases are now the default starting point for commercial post-training — the value moves to the fine-tune, not the checkpoint. (Unite.AI)

EU AI Act Article 50 guidance expands enterprise obligations

New guidance on the EU AI Act's Article 50 transparency obligations broadens what enterprises must disclose about AI-generated content and system interactions, per Law.com's analysis. Why it matters: compliance teams running open models in production now have a concrete disclosure checklist to map against their inference stack. (Law.com)

Watching tomorrow

DeepSeek's September 14 cutover, when deepseek-v4-pro API requests start routing to V4.1-Flash at Flash prices — the first forced migration of paying traffic onto the new architecture.