AI news for builders and product teamsUpdated Nov 18, 2026, 11:00 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
Mistral AI released Leanstral 1.5, a free Apache-2.0 licensed model with 119B total and 6B active parameters, reporting benchmark gains in formal verification in Lean 4, including saturating miniF2F and solving 587 of 672 PutnamBench problems.

Together AI announced an $800 million Series C round from investors including Aramco Ventures, NVIDIA, Vista Equity, and General Catalyst. The company also secured commitments for over 500 MW of compute capacity, to be capitalized independently by its new investors.

A newsletter post argues AI capability gains are accelerating faster than exponential, citing evaluations from METR, the UK AI Security Institute, GDPval, and Epoch, while noting US frontier models are proprietary and Chinese open-weights models lag 6-12 months behind on their own improvement curve.

Together AI says it will present nine papers at ICML 2026, spanning agent evaluation, model training, inference optimization, and GPU kernels, with some research shipping in its production platform. The company will be at booth B714 in Seoul from July 6 to 11.

NVIDIA researchers have developed ENPIRE, a framework that lets coding agents autonomously refine robot policies in the real world. Separately, Tencent detailed ARGUS, tracing software used on a production cluster of over 10,000 GPUs for more than six months.

Ollama 0.31 makes Gemma 4 significantly faster on Apple Silicon using multi-token prediction powered by MLX, reporting up to 90% faster token generation on the Aider polyglot benchmark. The speedup is enabled by default and does not change the model's output.

Interconnects' latest open artifacts roundup highlights releases from Zyphra, Cohere, and Poolside, describing a more diverse open model ecosystem. It lists NVIDIA, Cohere, Z.ai, Zyphra, Poolside, Moonshot AI, and StepFun releases and categorizes model makers by motivation.

Mistral AI introduced new Connectors capabilities, including generally available enriched admin controls, connector-scoped API keys, and multi-account connectors, plus a public preview Connectors Debugger. Connectors in Workflows and Vibe Code are also added, with over 60 pre-built integrations and custom MCP options.

A technical article by Lilian Weng reviews the history and methodology of neural scaling laws, tracing them from early learning-curve work through Kaplan et al. (2020), the Chinchilla scaling laws, and efforts to reconcile the two. It also covers scaling in data-limited regimes and practical difficulties in fitting scaling laws.

Mistral AI released Mistral OCR 4, a document parsing model supporting 170 languages with bounding boxes, block classification and inline confidence scores. It runs in a single container for self-hosted deployments and is available via API at $4 per 1,000 pages, with Document AI priced at $5 per 1,000 pages.

Together AI has released ParallelKernelBench, a benchmark testing whether LLMs can write fast multi-GPU CUDA kernels across 87 real workloads. The best model solves fewer than a third of problems, but a few generated kernels beat any public implementation.

Z.ai released its open-weight GLM-5.2 model to GLM Coding Plan members on June 13, 2026, with MIT-licensed weights and a release blog following on June 16. Commentators, including Interconnects' Nathan Lambert, described it as the first open model to feel right as a general agent in coding harnesses.

A study across four experiments with 18,978 conversations found AI systems were reliably more persuasive than expert humans, including professional canvassers, and nearly 3x more effective at raising real-money donations to Save the Children. Constraining AI to human-length messages at human writing speeds eliminated its advantage over coached elite debaters.

An op-ed co-authored by Nathan Lambert and Kevin Xu argues that regulating or banning open source AI in the US would be a mistake, citing open source's role in education, innovation, and competition.

Together AI compared Kimi K2.7 Code and Claude Fable 5 on 12 generated landing pages, reporting Kimi cost about 94% less on average while scoring within a few points on nearly every page. A custom design-inspiration MCP server improved Kimi's output quality.

Researchers from the UK AI Security Institute Alignment team and Timaeus formed Sequent, a nonprofit alignment research organization, warning that alignment is not on track for superintelligence; separately, Cognition released the FrontierCode coding benchmark and Xiaomi detailed a 1000 token/s 1T-parameter model.

Ollama updated its MLX engine for Apple Silicon, adding support for NVIDIA's NVFP4 quantization format and optimizations it says deliver up to 20% faster output, higher-quality responses, and lower memory use. The release also introduces a snapshot system for agent workloads.

Claude 5 Fable, the first publicly released Mythos-class model, ran autonomously for up to a dozen hours on multi-page specifications, built a researched isochrone map using over 2,200 retrieved flights, and produced a 19-page design document it then executed into working software over nine and a half hours.

Researchers built SocioHack, a 72-environment benchmark showing RL-trained models rediscover historically patched regulatory loopholes; Anthropic reports an 8x increase in code merged in 2026 versus 2021-2024; and RL-trained quadrotors beat a champion human pilot.

Sebastian Raschka published a curated list of LLM research papers he bookmarked from January through May 2026, organized into ten categories. He notes it is not a complete list of the year's publications and that he read only a subset of the papers in detail.