MiniMax launches M3.1 Flash as a cheaper agent model, with weights still to come
MiniMax released a low-cost hosted agent model, while its checkpoint date, license and local hardware requirements remain unconfirmed.
Inception launches Mercury 3, betting parallel text refinement can cut agent waiting time
Inception released Mercury 3 through hosted interfaces and reports 1,068 tokens per second, with independent task-level validation still pending.
doubleAI introduces Maxwell to rewrite running programs, with proof claims still under scrutiny
doubleAI says Maxwell replaces running binary code with faster equivalents and reports a 4K video pipeline falling from 250 to 134 seconds.
Tenstorrent raises $1.5 billion to expand manufacturing and Galaxy deployment
Tenstorrent closed a $1.5 billion Series E at a $6.2 billion valuation to expand manufacturing, systems deployment and chip-design intellectual property.
WhatsApp help pages carried hidden instructions telling a bot not to recommend regulators
Redact found machine-facing instructions on WhatsApp help pages, including a rule against recommending regulators, without demonstrating that outside chatbots obeyed them.
NorForge finds exposed Cairn dashboards showing patient, coordinated AI intrusion work
NorForge documented more than 20 likely malicious Cairn panels and a cluster with over 1,100 projects, while validating a live uploader shell.
DHS DeepSeek order reaches third-party implementations, leaving a model-provenance problem
A DHS notice excludes DeepSeek products from civilian-agency procurement and extends removal to third-party products developed using DeepSeek implementations.
CERN and Mistral outline a shared collider-physics foundation model
CERN and Mistral signed a statement of intent to develop a pretrained model from simulated collider events, without announcing funding or a delivery schedule.
MemT proposes persistent memory inside model computation, with a 33-fold state reduction in one test
MemT researchers report much smaller persistent conversation state, while keeping model weights fixed during interaction and trading away exact history access.
whatmodel opens a comparison of lab benchmark claims and independent scores
whatmodel reports a 6.4% average relative gap across 611 matched benchmark pairs, while warning that settings and selection limit the conclusion.