Daily Summary of AI · Live edition, updated continuouslySaturday 21 November 2026 · UTC

Latest AI News Today: 21 November 2026

Thinking Machines releases Inkling 2 Mini as Bytedance questions training phases

Thinking Machines released Inkling 2 Mini, an open looped transformer that runs locally on Macs and leads open models under 100 billion parameters on agentic coding tests. Bytedance and US universities argued that most AI training phases are unnecessary, while Baseten released decision models that top a decision index.

Dell said AI readiness depends on more than servers, storage and capacity, especially for financial institutions, with autonomous agents and token budgets in use.

Illustration of the day's news: Thinking Machines releases Inkling 2 Mini as Bytedance questions training phasesIllustration
DS
Daily Summary Of Editors · How we summarise
5 stories5 sources
Image from Illustration: Thinking Machines releases Inkling 2 Mini, a looped transformer for laptopsIllustration
01First reported

Thinking Machines releases Inkling 2 Mini, a looped transformer for laptops

Thinking Machines released Inkling 2 Mini, an open looped transformer that runs 12 of its 24 layers three times per token to reason internally with fewer output tokens. The 26.4 billion parameter model processes text and images, has a 1-million-token context window, and is licensed under Apache 2.0. It leads open models under 100 billion parameters on agentic coding tests and runs locally on Macs with at least 32 GB of memory.

Image from Illustration: Bytedance argues most AI training phases are unnecessaryIllustration
02First reported

Bytedance argues most AI training phases are unnecessary

A paper from Bytedance and several US universities argues that most training phases used to build AI agents are unnecessary, saying three steps are enough for web research, code execution and multi-step tasks. The paper says nested pipelines make it hard to identify helpful steps and raise reproduction costs. Bytedance supports the claim with a competitive open model.

Image from Illustration: Robot hands remain a barrier to replacing human workersIllustration
03First reported

Robot hands remain a barrier to replacing human workers

Experts say dexterity is a key barrier to AI replacing human workers, because human hands can perform many useful actions that robots still struggle to match. In an Amazon warehouse near Luton, computers direct where items are stored, how boxes are packed and how much tape is used, while humans provide the hands. Humanoid robots have made advances, but replicating the breadth of the human hand remains one of the biggest challenges in robotics.

Image from Illustration: Baseten releases Bison-1 and Groundhog-1, topping decision model indexIllustration
04First reported

Baseten releases Bison-1 and Groundhog-1, topping decision model index

Baseten released bison-1 and groundhog-1, open-weight decision models built on Google's Gemma 4 that return calibrated probabilities instead of text. Bison-1 scores 76 on the Artificial Analysis Decision Index v2.1, ranking first of 19, while both models raise context from 16k to 128k tokens and accept images. The models target routing, triage and guardrail pipelines, where longer context can fit full policies and histories in one call.

Image from Illustration: Dell details AI agents, token budgets and threat landscape at New York SymposiumIllustration
05First reported

Dell details AI agents, token budgets and threat landscape at New York Symposium

At Dell's New York Symposium, executives said AI readiness depends on more than servers, storage and capacity, especially for financial institutions facing reliability, regulation, returns and a changing threat landscape. Dell has run fully autonomous agents internally for nearly two years, storing knowledge in machine-readable skill files, while finance teams set token budgets and weigh on-premises workloads.