DeAI Daily Brief — 7 November 2026

Today in DeAI: Inkling 2 Small claims the western open-weight lead while Ripple-65B Preview ships Apache 2.0 weights now, Nous Research publishes Arete-27B, and Baidu's ERNIE license excludes the EU by name.

Ripple-65B ships Apache 2.0 weights while Inkling 2 Small waits for its repo

Two American open-weight releases bookend the licensing question. Ricursive Intelligence's Ripple-65B Preview — a 65-billion-parameter dense model, 12 trillion training tokens, 1M context — has been downloadable under Apache 2.0 since November 4, and the company claims 80.52 on TerminalBench 2.1 at a stated $32 million total R&D cost. Thinking Machines Lab's Inkling 2 Small (594B total, 17B active) claims a better Artificial Analysis Intelligence Index score of 31, but its weights remain Tinker-only pending the promised Apache 2.0 drop; no repo exists yet. All benchmark figures are provider self-reports. Why it matters: the model you can legally download and serve beats the model with better claims whenever the decision is "what do I run this week." (The Decoder, Hugging Face) — our full coverage

Nous Research publishes Arete-27B with weights, evals, and training data

Nous Research released Arete-27B on Hugging Face the evening of November 6, a 27-billion-parameter fine-tune of Qwen3.8-27B that the lab describes as aligned to Western values. The release ships weights, evaluations, and training data, and community GGUF and EXL3 quantizations appeared within hours — bartowski's first, about two hours after the drop. Why it matters: a fully open drop on single-node 27B hardware makes local fine-tuning and reproducible evals practical the day it ships; the alignment positioning is the lab's own description until independent evals run. (Hugging Face)

Baidu's ERNIE 5.5 Flash license defines its Territory as excluding the EU, UK, and South Korea

The LICENSE file on Baidu's ERNIE 5.5 Flash Base repo — live since October 21 — defines "Territory" as worldwide other than the European Union, the United Kingdom, and South Korea. The restriction is in the license text itself, verified by reading the file. Why it matters: open weights and deployment freedom are different products; a team in Berlin or London can download the weights but cannot lawfully serve them under this license, which matters for EU data-center inference and multi-region products. (Hugging Face)

Revolut claims 10x coding-inference savings from self-hosting, methodology unpublished

Revolut says it cut AI coding inference costs roughly 10x in under a year by moving off Anthropic's API to self-hosted models. The Register reports the figure as a company self-report with no published methodology — no model list, no hardware spec, no token volumes, no quality comparison against the Claude baseline it replaced. Why it matters: a 10x raw-cost claim can coexist with worse per-task economics if acceptance rates drop, so treat it as one data point in the self-hosting ledger, not a verdict. (The Register)

GitLab patches critical AI Gateway RCE, CVE-2026-90970

GitLab released urgent fixes for a critical vulnerability in its AI Gateway that lets authenticated attackers achieve remote command execution. Teams running self-hosted GitLab AI Gateway instances should patch immediately and rotate the provider credentials the gateway holds; managed-instance operators should confirm their fixed version. Why it matters: an AI gateway sits in the token path with provider keys, so an RCE there is credential-and-log exposure, not just downtime. (Cyber Security News)

Watching tomorrow

The Inkling 2 Small Apache 2.0 repo — Thinking Machines said "next week" on November 5, and the Hugging Face drop is the moment its benchmark claims become independently testable. Qwen's "The becoming begins" teaser also has a weekend-drop rumor attached, though no Qwen 4 repo exists yet.