EN · International edition · 14/11/2026
Inferama daily · 14/11/2026
The most important AI developments since the previous newsletter, selected for editorial relevance.
News coverage period: 13/11/2026 08:30 → 14/11/2026 08:30 (Europe/Madrid).
Runway releases Praxis-1 SO-101, a video-and-action policy for LeRobot arms
Dated November 12, 2026, the checkpoint controls SO-100/SO-101 arms through joint-space actions. The reference version occupies 54.4 GB and calls for an 80 GB GPU; its NVFP4 variant reduces inference memory experimentally but downloads the same complete weights.
Read article →
Google launches Gemini 4 Hydrogen with global availability and per-token pricing
Gemini 4 Hydrogen was introduced on November 12, 2026, for everyday coding and agentic workflows. Google announced a global rollout, a unified one-million-token context window, and standard API rates of $1.50 per million input tokens and $7.50 per million output tokens.
Read article →
Liquid AI releases LiquidCodec, a real-time speech codec
LiquidCodec encodes mono speech as discrete tokens and reconstructs audio from them. The official repository provides BF16 weights; the speed and quality figures come from Liquid AI’s tests, not an independent reproduction.
Read article →
Alibaba Cloud opens Qwen4-Max-Preview preview exclusively via Token Plan
Qwen4-Max-Preview was available on Alibaba Cloud from November 11 to 25, 2026, exclusively for Token Plan subscribers. The documentation describes text, image, and video input, reasoning, and tool calling; it also sets a shared output limit and confirms the model will be retired when the preview ends.
Read article →
Generative AI in the classroom: how to use it without losing sight of learning
A practical guide to deciding when to allow, limit or incorporate generative AI in secondary and university teaching. Learn how to redesign assessments, teach AI literacy, protect student privacy, set disclosure rules and understand the limits of AI detectors.
Read article →
Semantic search: definition, workflow, examples, and limits
Semantic search retrieves information by estimating meaning similarity through vector representations, rather than relying only on exact term matching. This guide explains the technical flow, its differences from lexical search, and criteria for evaluating performance.
Read article →
Qwen3-30B-A3B MoE-Output SAE: what it analyzes and what it does not prove
ByteDance Seed released three sparse autoencoders for studying the MoE block outputs of Qwen3-30B-A3B. The paper describes expert specialization in three layers, but its findings apply to a specific study and do not establish causality or general model performance.
Read article →
How to Recover AI Agents After a Failure Without Duplicating Effects
An interrupted agent may have completed an external action even if it never saved the result. This guide explains how to persist state, classify failures, retry with limits, reconcile uncertain operations, and escalate to a person when repeating an action is unsafe.
Read article →
Anthropic explains how Claude’s text watermark works and what it can detect
Anthropic describes a statistical watermark designed to indicate when Claude may have contributed to a text. The signal does not prove exclusive authorship, may be hard to detect in short passages, and its detector remains in private preview.
Read article →
Google launches Gemini 4 Hydrogen, a faster model for AI agents and coding
Google announced Gemini 4 Hydrogen on November 12, 2026, positioning it for fast agent workflows and everyday coding. The model is rolling out through Gemini, Google Search and developer tools, with API pricing of $1.50 per million input tokens and $7.50 per million output tokens. Google says it completes tasks 2.5 times faster than Gemini 3.8 Flash under its default reasoning setting; that speed claim is tied to the company’s comparison, not an independently replicated result.
Read article →