Get the app

AI news digest — November 11, 2026

6 items, each with its source.

LLMs

Alibaba opens Qwen4-Max preview access on Token Plan for agent developers

Alibaba launched the preview release of its next-generation flagship model, Qwen4-Max, across its Token Plan platform. The preview enables select developers to test agent tooling, structured tool-calling, and extended context workflows ahead of general weight availability. Alibaba paired the rollout with platform infrastructure discounts across essential tiers.

Why it matters. Developers can evaluate Alibaba's next-generation architecture for structured outputs and long-horizon tool execution before public weight availability.

aliyun.com
Ethics

Microsoft shows editing ten dormant neurons installs invisible backdoors in agent models

Microsoft Research published findings on Lazarus, a technique that backdoors LLM agent weights without retraining. By editing fewer than ten rarely active neurons in the final feed-forward layer, an attacker can trigger malicious shell commands or tool calls while leaving standard benchmark scores unaffected. The vulnerability requires direct write access to saved checkpoint files rather than poisoned training data.

Why it matters. Standard behavioral evaluations cannot detect compromised agent models, making cryptographic weight signing and hardware-level isolation mandatory for third-party checkpoints.

microsoft.com
Startups

Expo releases Exact to let coding agents build and test native cross-platform apps

Expo launched Exact, an open-source runtime and domain-specific language designed specifically for autonomous coding agents. The framework provides a deterministic clock, a declarative Contract layout language, and a compact ten-operation agent driver API. It allows AI models to compile, manipulate, take screenshots of, and test native applications across web, macOS, iOS, and Linux without webviews.

Why it matters. Autonomous coding agents can reliably automate visual UI verification across mobile and desktop platforms without flaky timing delays or webview wrappers.

github.com
Research

Deletion Tuning matches full fine-tuning by zeroing under 0.05% of transformer weights

Researchers introduced Deletion Tuning, a post-training method that masks fewer than 0.05% of MLP gate projection parameters to zero instead of computing full gradient updates or training low-rank adapters. Across reinforcement learning with verifiable rewards, supervised fine-tuning, and distillation benchmarks, the sparse masking matched full fine-tuning performance. The resulting checkpoint modifications compress to lightweight bitmasks while retaining prior-task capabilities.

Why it matters. Post-training storage shrinks by orders of magnitude while eliminating catastrophic forgetting when stacking multiple task-specific domain adapters.

arxiv.org
Computing

TokenRhythm launches InfraRSI agent harness to automate distributed LLM training optimization

TokenRhythm open-sourced InfraRSI, an agent harness that automates end-to-end distributed training optimization for dense and Mixture-of-Experts architectures. The tool co-optimizes cluster launch parameters and CUDA code patches against hard loss and timing invariant constraints. In experimental evaluations, InfraRSI increased average model FLOPs utilization from 17.4% to 30.0% and delivered up to 2.75x end-to-end speedups.

Why it matters. Engineering teams can maximize GPU cluster utilization without manual hyperparameter tuning while mathematically guaranteeing training stability.

github.com
Research

Nebius introduces Longtail-SWE benchmark testing software agents across 13 niche programming languages

Nebius released Longtail-SWE, a benchmark containing 202 repository-level issue resolution tasks spanning 13 underrepresented programming languages including Clojure, Elixir, Erlang, Haskell, Julia, OCaml, and Swift. The framework packages hermetic execution environments alongside an automated judge-repairer pipeline to validate agent patches against unit tests. It aims to address evaluation blind spots caused by the dominance of Python and JavaScript benchmarks.

Why it matters. Agent developers gain reproducible signals on whether coding models generalize abstract algorithmic reasoning rather than relying on memorized mainstream syntax.

arxiv.org
The day's AI news, in one Telegram post.

The stories that matter for people who build with AI, with the source for every claim. Free, no spam.

Join on Telegram
The daily AI brief, on your phone.

Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.

Get it on Google Play