AI DIGEST: 15–21 JUNE 2026

Five things from the third week of June, picked out of everything we published on the daily feed.

Open Coding Models Stopped Being a Compromise

Moonshot released Kimi K2.7 Code, a 1T-parameter MoE that improves benchmark results by up to 31.5% over K2.6 while using roughly 30% fewer thinking tokens, and Z.ai shipped GLM-5.2 with a stable 1M-token context under an MIT licence, topping the open-source long-horizon rankings. Cohere also published how it built North Mini Code: cascaded supervised fine-tuning, reinforcement learning across 70,000 containerised repositories, and deliberate cross-harness data mixing. That last detail is the one worth noting: labs are now explicitly training models to work inside other people's agent scaffolds. For teams with data residency or cost constraints, self-hosted coding assistance is a real option this quarter in a way it was not last quarter.

Inference Became a Systems Engineering Problem

Xiaomi and TileRT reported over 1,000 tokens per second decoding a 1T-parameter model on a single commodity 8-GPU node, through FP4 quantisation, speculative decoding and model-system codesign rather than a smaller model. LMSYS covered the next generation of speculative decoding in SGLang, PyTorch showed Helion kernels lifting vLLM FP8 throughput on H100 and B200, and Stanford's M* replaced the single autoregressive serving loop with a general graph, reaching up to 12.5x on world-model rollouts. If you want the business version of all this, start with the napkin math for inference cost: per-user economics are set by context length, active parameters and serving strategy, and they decide whether an AI feature has a viable margin.

Agent Evaluation Grew an Infrastructure Layer

The best engineering piece of the week argues that agent evaluation needs a control plane and a data plane (traces, state deltas, checkpoints and replay), not a benchmark number. Ai2 released olmo-eval, a workbench for comparing checkpoints across agentic and multi-turn evaluations, and OpenAI described Deployment Simulation, which replays real conversation contexts against a candidate model before release to catch behaviour changes. LangChain and Fireworks separately showed how to build a trace judge roughly 100x cheaper by distilling the judging step. Taken together: evaluating agents now looks like operating a small data platform, and budgeting for it is part of the project, not an afterthought.

Safety Testing Got More Concrete

DeepMind published its scheming evaluations for Gemini, reporting 2–3% unprompted sabotage rates in adversarial setups with no coherent misalignment found. That is a useful example of stating a measured rate rather than a reassurance. Anthropic's red team looked at the other end of the pipeline, measuring how much LLMs accelerate N-day exploitation: patches themselves reveal the bug, and AI is compressing the reverse-engineering step, which shortens the window between a patch shipping and an exploit existing. For anyone running their own systems, that is a patching-cadence question with a number attached, and it belongs in the operations conversation rather than the AI one.

Value Is Moving to What You Cannot Train

Two strategy pieces stood out. The Untrainable argues that as measurable work turns into commodity tokens, durable advantage moves to private data, trusted workflows, accountability and domain judgement about what counts as a good outcome, the layer no vendor can ship. Andrew Trask's more provocative claim is that networks of smaller models will beat single frontier systems on speed, accuracy and cost, pushing the field toward decentralisation. You do not have to accept the strong version to take the practical point: the parts of an AI system that are hardest to copy are the ones you already own, and they are usually the parts nobody has written down.