EN · International edition · 15/11/2026
Inferama daily · 15/11/2026
The most important AI developments since the previous newsletter, selected for editorial relevance.
News coverage period: 14/11/2026 08:30 → 15/11/2026 08:30 (Europe/Madrid).
Deep Principle releases Hydrogen 1.0 in INT8 format for download
Deep Principle has published INT8 weights for Hydrogen-1.0, a quantized version of its scientific model based on GLM-5.3. The repository provides the files under the MIT License, but does not confirm a hosted API or the original publication date. The checkpoint is about 793 GB, and the team reports testing it with SGLang on eight NVIDIA B300 GPUs.
Read article →
Data contamination in AI: what it means and how to interpret its effects
Data contamination occurs when evaluation material also appears in a model’s training data. It can complicate how a benchmark score should be read, but a match alone does not prove memorization or show how much a score changed.
Read article →
Embeddings: how vector representations are learned and what their geometry can measure
An embedding turns a data unit—such as a word, document or image—into a numerical vector. Its geometry can help compare items according to what a model learned, but proximity depends on the objective, data and metric; it does not prove that two items are identical or share a universal meaning.
Read article →
Praxis-1 YAM: capabilities, requirements, and limits for controlling two robotic arms
Praxis-1 YAM combines video prediction and actions to control two YAM arms. The published checkpoint requires an 80 GB GPU in its reference configuration; the NVFP4 option reduces memory on certain GPUs but is experimental and does not equate to a real-world control evaluation.
Read article →
ManiFlow-tb: What It Controls, What Its Results Measure, and What It Takes to Run
ManiFlow-tb is a low-level policy for manipulating blocks with a simulated Franka arm. Its published ToyBlocks evaluation measures execution under a reference plan—not a complete robot system or performance on arbitrary robots.
Read article →
Molmo2-4B-tb: capabilities, limits, and deployment costs
An analysis of ToyBlocks’ high-level planner: the block-stacking tasks it targets, the reported simulation and real-robot results, the hardware and licensing considerations, and when this specialized model may fit a robotics project.
Read article →
API contracts for AI systems: schemas, versioning, and resilient integrations
A clear API contract helps an AI integration handle format changes, errors, and service variations without turning every unexpected response into a downstream failure. Learn how to define and validate interfaces, evolve them compatibly, and test recovery—without mistaking structural validation for proof that an AI output is correct.
Read article →
OpenAI launches ChatGPT Wallet for AI agents to make purchases with user approval
The new digital wallet allows OpenAI's AI agents to complete transactions on behalf of users, subject to biometric verification and card tokenization. Currently, it is only available for Pro subscribers in the United States.
Read article →
AesCode Infographics Style: What the aesthetic dimension measures and how to interpret its score
The Style metric in AesCode evaluates perceived visual quality using a sample-specific rubric, separating it from factual correctness and code functionality. This analysis details its methodology, limits, and conditions for valid comparison.
Read article →
OpenFold3 arrives on SageMaker JumpStart to predict biomolecular complex structures
AWS confirms the availability of OpenFold3 on SageMaker JumpStart, enabling teams to predict structures of proteins, DNA, RNA, and ligands through asynchronous inference and multi-GPU parallelism.
Read article →