DeAI Daily Brief — 10 November 2026

Today in DeAI: NEAR AI pauses Cloud inference after reporting unauthorized access to its internal systems, a Qwen 27B variant ships weight edits made with no training at all, and a patched llama.cpp stack runs a 284B-parameter model on 64GB of RAM.

NEAR AI reports unauthorized access, pauses Cloud inference

NEAR AI said on November 9 that it identified unauthorized access to internal systems and paused NEAR AI Cloud inference as a precaution, with its own status page showing Chat API at 92.59% uptime for the week. The company says customer workloads stayed inside hardware-isolated enclaves and that NEAR Protocol was not affected; it has not published a postmortem. Why it matters: a confidential-inference provider going dark is a reminder that hardware isolation protects the prompt path, not the operator plane — availability belongs in the private-inference decision. (NEAR AI status page, Crypto Briefing) Full story

Qwen Concept 27B rewrites weights with no training run

A team behind the ConceptCountry organization released Qwen Concept 27B, a Qwen3.8-27B variant modified by rank-one additive edits compiled into nine down-projection weight matrices — no gradient steps, no training data. The team's research page reports GPQA Diamond up 4.3 points, LiveCodeBench v6 up 13.0, and thinking-token output down 27–54%, with a norm-matched random-edit control whose confidence intervals straddle zero. Weights are on Hugging Face in bf16 and NVFP4, under an Apache-2.0-linked license. Why it matters: if the self-reported deltas replicate, post-release weight editing becomes a cheap way to cut thinking-token spend on models self-hosters already run. (Hugging Face, concept.country)

llama-modchip runs a 284B MoE on 64GB of RAM, no GPU

The synchopate project's patched llama.cpp stack runs DeepSeek-V4-Flash's 284B-parameter MoE (about 13B active) at roughly 9–10 tokens per second generation on a 64GB desktop, using ~1.71 bits per weight — numbers the project reports and Hacker News commenters are already stress-testing at 402 points. Why it matters: the hardware floor for frontier-scale open weights keeps dropping, though quality at 1.7 bits is unverified and prefill is slow for agentic loops. (GitHub)

Meta prices Muse Ember at $0.10 per million input tokens

Meta released Muse Ember, its smallest model yet, pitched at the tool calls and extraction loops agents repeat all day, with TechCrunch reporting a claimed $0.10 per million input tokens and Meta's announcement calling it built for agentic workloads like tool calls. Artificial Analysis scores it 36 on its Intelligence Index and calls it the cheapest model at that score. Whether weights are open remains unconfirmed — no weights repo was located at publish time. Why it matters: if the price holds, the boring 90% of agent work gets priced near zero, but the reported coding weakness means it is not a drop-in coder. (TechCrunch, Meta (@alexandr_wang))

Anthropic gates Mythos 5.2 behind verified organizations

Anthropic shipped Claude Fable 5.2 to everyone and locked Mythos 5.2 — the cyber-capable tier, priced at $8/$40 per million tokens with 1M context — behind its verified-organization programs, per its platform docs. Why it matters: identity verification is becoming an access-control layer for the most capable models, and gated tiers do not reach open weights. (Anthropic platform docs)

Watching tomorrow

Whether NEAR AI publishes a durable incident postmortem, and whether Meta confirms or denies open weights for Muse Ember.