MiniMax M3 1M context, Google Gemma 4 12B, Ideogram open weights

MiniMax released its M3 model, featuring a 1‑million‑token context window and native multimodal input, initially available via API and MiniMax Code. The company says the model’s weights and a technical report will be posted on Hugging Face and GitHub within the next ten days.

MiniMax introduced MiniMax Sparse Attention (MSA), a new sparse attention architecture that reduces quadratic complexity and lets its M3 model handle up to 1 million tokens. The redesign cuts per‑token compute to a twentieth of its predecessor, delivering 9× faster prefilling and 15× faster decoding while preserving capability.

Google’s DeepMind team released Gemma 4 12B, a 12‑billion‑parameter encoder‑free multimodal model that handles text, images, and audio. Its compact footprint lets it run locally on consumer laptops with 16 GB memory, enabling agentic workflows without cloud APIs.

Ideogram 4, the first open-weight text‑to‑image model from Ideogram, offers 9.3 B parameters in NF4 and FP8 quantizations and supports 2k resolution with structured JSON prompts. Benchmarks show it leads all open‑weight models on design‑focused and general image generation tasks. The code and weights are released on GitHub and HuggingFace.

TripoSplat, an open‑source model from VAST‑AI, converts a single 2D image into a variable number of high‑quality 3D Gaussians, enabling asset creation for AR/VR, games, and simulations. The weights are freely available, can run locally, and integrate directly with ComfyUI for rapid experimentation.

The Alignment Research Center (ARC) has partnered with AIcrowd to host the White-Box Estimation Challenge, a competition focused on developing better estimation algorithms for AI alignment and interpretability. Participants will submit models that infer hidden parameters from transparent environments, aiming to advance safety research tools.

Researchers measured the interaction between reasoning and truthfulness across 63 language models and found a phase transition around 3.5 billion parameters. Below this size the two capabilities anticorrelate, while above it they cooperate, and the transition can be shifted by architecture, data curation, and training tricks. The paper provides a dashboard and predictive tools.

The paper introduces Discounted Reward for Same‑Length Trajectories (DReST), a reward shaping that trains agents to be neutral about shutdown timing while remaining useful. Applied to deep RL agents and instruction‑tuned LLMs, DReST improves shutdown resistance and generalizes to unseen contexts, halving the likelihood of agents influencing shutdown.

GitHub’s move to token‑based billing for Copilot signals the end of subsidized US AI coding tools, exposing the 10× premium prices of frontier models. Cheaper alternatives like Kimi and DeepSeek are eroding US market share while raising data‑privacy and pricing concerns.

Uber announced it will cap token spending on each AI coding assistant, such as Claude Code and Cursor, at $1,500 per employee per month after blowing its 2026 AI budget in four months. The move seeks to curb runaway costs while still letting engineers benefit from AI tools.

Perplexity announced a new hybrid inference orchestrator that decides per request whether to run on‑device models for lightweight or privacy‑sensitive tasks and cloud models for heavy reasoning. The system, part of its Personal Compute initiative, aims to maximize token‑per‑watt efficiency while preserving privacy and reducing server load.

Anthropic announced that it is extending its Project Glasswing partnership to about 150 new organizations across over 15 countries, adding critical‑infrastructure players in sectors like power, water, healthcare, communications and hardware. Each new partner must meet security requirements before gaining access to the Claude Mythos Preview model, which has already identified more than 10,000 high‑severity flaws.

President Donald Trump issued an executive order that creates a voluntary framework requiring AI developers to give the federal government up to 30 days of early access to ‘covered frontier models’ for cybersecurity review before public release. The order also directs agencies to establish a classified benchmarking process to identify such models.

NeurIPS 2026’s Position Paper Track applied an AI‑detection system to enforce its new policy that papers be substantially human‑written. The detector led to the desk‑rejection of 178 submissions (18.4% of the track) and requests for proof of human involvement from another 123 papers, sparking community concerns about calibration and fairness.

Wall Attention is a new attention variant that adds per‑channel, per‑timestep decay via persistent 'wall' memory tokens, enabling better long‑context reasoning while keeping compute low. The open‑source implementation provides fused training kernels and a fast decode kernel, supporting GQA and scalable inference.