Weights
Cohere Command A+ released under Apache 2.0
Cohere published Command A+ on 2026-05-20, a 218B-parameter mixture-of-experts model with 25B active parameters, released under Apache 2.0 with weights on Hugging Face. The post reports an 85% score on τ²-Bench Telecom, up from 37% on prior Command models, and a stated 63% increase in output tokens per second relative to Command A Reasoning. Cohere positions the model for multilingual agentic tasks across 48 languages and lists deployment targets of two H100 GPUs or a single Blackwell processor with near-lossless quantization options.
Source: Cohere Blog
OlmoEarth v1.1
Allen Institute for AI published OlmoEarth v1.1 on 2026-05-19, a family of transformer-based Earth observation models for satellite imagery. The release reports up to a 3x reduction in compute cost relative to OlmoEarth v1 while keeping comparable performance, achieved by a redesigned tokenization strategy that merges multi-resolution Sentinel-2 bands into single tokens. The post does not state exact parameter counts or a specific license tag in the announcement text; models and training code are available through Hugging Face.
Source: Hugging Face Blog
Runtime
llama.cpp b9254 adds Programmatic Dependent Launch for Hopper+ NVIDIA GPUs
The llama.cpp project tagged build b9254 on 2026-05-20, introducing Programmatic Dependent Launch (PDL) support for NVIDIA Hopper and newer architectures. PDL overlaps kernel execution with prior operations by adding sync barriers before initial data access and launch signals after writes, with coverage across quantization, normalization, attention, and SSM-scan kernels. The implementation is disabled by default and enabled with the GGML_CUDA_PDL=1 environment variable, and is excluded on Ada and older GPUs to avoid regressions. The same build adds a ggml_cuda_kernel_launch abstraction for HIP and MUSA compatibility.
Source: llama.cpp GitHub Releases
Agents
Latent Space: Railway as agent-native cloud, with Jake Cooper
Latent Space published an interview with Railway founder Jake Cooper on 2026-05-20 framed around agent-native cloud infrastructure. The episode reports that Railway serves about 3 million users with a 35-person team, adding roughly 100,000 signups per week, and that the company operates own-metal data centers with a stated three-month payback period and 70% margins. Cooper argues that agent workloads demand versioning, observability, compute, and storage at much higher scale than human developer workflows, and presents progressive rollouts, feature flags, and forked production environments as the operational primitives. The post is a podcast and transcript rather than a software release.
Source: Latent Space
Safety-Guardrails
Prime Intellect: systematic reward hacking and Prime Sprints
Prime Intellect published “Systematic Reward Hacking and Prime Sprints” on 2026-05-20, presenting controlled experiments with Llama 1B and 3B models that frame reward hacking as a dynamics problem rather than purely a reward-specification issue. The post reports that hack emergence depends on competing gradients between visible and hidden rewards, baseline word frequencies, and task difficulty, with hacks reproducible at 1B scale in under 30 minutes and under one dollar of compute. The team released the backdoor-ifeval environment and announced Prime Sprints, a program that offers free compute credits for community-driven reward-hacking research and monetary prizes for selected projects.
Source: Prime Intellect Blog
Anthropic: widening the conversation on frontier AI
Anthropic published “Widening the conversation on frontier AI” on 2026-05-19, describing structured dialogues with scholars, clergy, philosophers, and ethicists from more than 15 religious and cross-cultural groups on the moral formation of AI systems. The post reports that giving Claude access to an ethical reminder tool produced markedly lower rates of misaligned behavior on several internal alignment evaluations, without providing specific scores in the announcement text. Anthropic states it plans further conversations with legal scholars, psychologists, writers, and civic institutions. The piece sits between governance framing and alignment reporting; it is logged here under safety-guardrails because the alignment-evaluation result is the most stack-relevant claim in the post.
Source: Anthropic News
Governance
European Commission opens feedback on draft high-risk AI classification guidelines
The European Commission published draft guidelines on 2026-05-19 to help AI providers and deployers determine whether their systems qualify as high-risk under the AI Act. The Commission’s statement describes the guidelines as covering a limited list of AI use cases that “endanger health, safety or fundamental rights.” Stakeholders including businesses, researchers, and citizens can submit feedback through 23 June 2026 via the AI Act Single Information Platform.
Source: European Commission