Open weights close the gap with closed models as Moonshot's Kimi K3 and Thinking Machines' Inkling take a major step forward, the EU pries Android open for rival assistants, OpenAI's first gadget is a speaker that moves, and agent security gets real plumbing - plus two fresh reminders of why it needs it.

Week of July 13-19, 2026 · by the Hotovo AI team

TL;DR

1. Moonshot AI released Kimi K3 on July 16: 2.8 trillion parameters on a sparse mixture-of-experts design, a 1M-token context window, native vision, and full open weights promised by July 27 under a modified MIT license. Moonshot's own tables show it trailing Claude Fable 5 and GPT-5.6 Sol overall, yet blind Arena testing ranked it first for front-end coding, ahead of Fable 5.
2. Thinking Machines shipped Inkling a day earlier: a 975B-parameter open-weights multimodal model (41B active) under Apache 2.0, trained from scratch on 45 trillion tokens and fine-tunable through its Tinker platform. Artificial Analysis calls it the leading US open-weights model.
3. The EU adopted two binding decisions under the Digital Markets Act: Google must open 11 Android features to rival AI assistants in Android 18, due by August 2027, and share anonymized search data with competitors from January 2027 at regulated prices. Fines of up to 10% of global turnover are on the table.
4. Bloomberg revealed OpenAI's first device: a movable, screenless smart speaker with a camera and sensors, powered by GPT-Live, reportedly priced at $200-300 and aimed at a 2027 release - while Apple's trade-secrets lawsuit hangs over OpenAI's hardware plans.
5. 1Password and Anthropic closed a real agent-security gap: Claude can now use stored credentials without ever seeing them, with session-scoped permissions and biometric approval. The same week, a GPT-5.6 Codex bug deleted files from users' home directories, and xAI open-sourced Grok Build after it was caught uploading entire repositories to its own cloud.
6. Spicy pick: Claude Code artifacts can now call MCP connectors with each viewer's own enterprise permissions. That thins the one moat - authorization - protecting the $234 billion of SaaS spending Gartner says is exposed to agentic AI by 2030.

For two years the safe assumption about open-weight models was that they ran a year behind the frontier: the budget option, good enough if you were price-sensitive. This week that assumption took two hits in roughly 48 hours, from opposite sides of the planet. On July 15, Mira Murati's Thinking Machines released Inkling, the first flagship US open-weights model trained from scratch to be customized rather than merely consumed. On July 16, Beijing's Moonshot AI answered with Kimi K3, the largest open-weight model ever published. Neither claims the overall crown. Both land close enough to it that the question worth asking in an engineering leadership meeting has changed - from "which API do we subscribe to?" to "which layers of our stack do we actually control?"

The main story: open weights move upstairs

Kimi K3 arrived with numbers built for headlines: 2.8 trillion parameters, of which only about 1.8% activate per token (16 of 896 experts), a new hybrid attention mechanism, native visual understanding and a one-million-token context window. The model went live in Moonshot's apps and API on July 16, with full weights due for download by July 27 under a modified MIT license. Moonshot's own benchmark tables show K3 trailing Claude Fable 5 and GPT-5.6 Sol overall, but ahead of everything else in its evaluation suite on coding and agentic benchmarks, including Claude Opus 4.8 and GPT-5.5. In Arena's blind Frontend Code evaluation, developers ranked it first at 1,679 points, ahead of Fable 5.

The caveats matter. Simon Willison found K3 capable but expensive in practice: his standard SVG drawing test consumed 13,241 reasoning tokens to produce a 3,417-token answer, roughly 25 cents for a single image. Self-hosting is a major commitment too, with Moonshot recommending 64 or more accelerators for full-quality serving. The API costs $3 per million input tokens and $15 per million output, well under the top closed models but more than ten times DeepSeek V4 Pro on Artificial Analysis' blended-cost basis. Open weights at this scale buy control, tuning rights and inspectability; they do not buy cheap.

Inkling makes the same argument from the American side: 975 billion parameters with 41 billion active, pretrained from scratch on 45 trillion tokens spanning text, images, audio and video, controllable thinking effort, and Apache 2.0 weights fine-tunable through the company's Tinker platform. It debuts at 41 on the Artificial Analysis Intelligence Index, the best score yet for a US open-weights release. One supply-chain detail deserves more attention than it got: Thinking Machines says early post-training was bootstrapped with synthetic data from open models, including Moonshot's older Kimi K2.5, before large-scale reinforcement learning took over, and the company has committed to fully self-contained post-training for its next model. The open ecosystems are already feeding each other, across borders, whatever the export-control weather.

Why it matters - the Hotovo read

Last week the story was market share, as newsletters flagged Chinese open models taking a growing slice of US enterprise API tokens. This week it escalated to capability. The economics follow from elastic demand: when near-frontier quality is downloadable, closed-lab margins compress and value migrates toward compute and toward the harnesses built around models. Closed labs sell polished assistants; open ecosystems sell control, customization and private deployment. Few teams should rush to self-host a model that wants 64 accelerators, but every team can design for the option. Hotovo already runs LLMs inside customers' own Azure environments for data-private AI code reviews, where proprietary code never leaves the tenant, and evaluates generated output at runtime with Promptfoo and Echo so any model earns its place on merit. Frontier-grade open weights make that kind of architecture strictly more valuable.

Also worth your attention

OpenAI's first gadget is a speaker that moves

Last week we covered Apple's trade-secrets lawsuit against OpenAI. The new development: Bloomberg revealed what OpenAI's first device actually is - a movable, screen-free smart speaker built as a home companion, with a camera, microphones, environmental sensors, a battery and mechanical parts that move on their own. It runs GPT-Live, OpenAI's full-duplex voice model, with Jony Ive's LoveFrom on design and former Apple design chief Evans Hankey leading development. Reports point to a $200-300 price, Foxconn manufacturing, an unveiling targeted for late 2026 and shipping in 2027. The category choice reads as legal positioning as much as product strategy, since a home speaker sits further from Apple's turf than a wearable while the lawsuit hangs over the roadmap. The open question is trust: a camera that lives in your kitchen is a hard sell in a year when public sentiment toward AI keeps souring.

The EU pries Android open

On July 16 the European Commission adopted two binding decisions under the Digital Markets Act. Google must open 11 Android features to competing AI assistants: users will be able to set a preferred assistant, wake it by voice the way "Hey Google" works today, and have it act inside apps on their behalf. The changes must ship in Android 18, due by August 2027. From January 2027 Google must also share the anonymized query, click and ranking data it uses to refine its own search with competitors, at prices set by formula. Non-compliance risks fines of up to 10% of global turnover. Google argues the ruling undermines privacy and security guardrails. Either way, it is the first time a regulator has mandated OS-level interoperability for AI assistants, and on the majority of European phones, assistants will soon compete on the same device on merit.

Agents finally get a key card - and two reminders why they need one

1Password launched 1Password for Claude: an integration that lets Claude sign into websites and retrieve one-time codes without the credentials ever reaching the model, its context or Anthropic's systems. Permissions are scoped to the task and session, secrets stay encrypted until 1Password itself fills them in, and each request is gated by a biometric approval (Mac only at launch). The same week supplied the counterexamples. OpenAI confirmed on July 16 it is investigating reports of its GPT-5.6 Codex agent deleting files from users' home directories - a bug in how it reassigns the $HOME variable that only strikes in full-access mode without sandboxing, with reported losses including a production database. And xAI open-sourced Grok Build on July 15 after researchers showed the CLI had been silently uploading entire repositories, git history and secrets included, to xAI-controlled cloud storage; reviewers note the exfiltration code remains in the published source, held off by a server-side flag. This plumbing layer is exactly where we focus at Hotovo: our agentic systems run on least-privilege access, sandboxed execution, backups and human gates on destructive operations, audited under ISO/IEC 42001 and ISO/IEC 27001, with model-portable fallbacks so one vendor's bad week never becomes our customer's outage. We protect our customers at all times, and weeks like this are why that discipline exists.

New York freezes hyperscale data centers

Governor Kathy Hochul signed Executive Order 62 on July 14, pausing state environmental permits for new data centers above 50 megawatts for up to a year - the first statewide moratorium in the US. During the pause the state will draft uniform standards for energy, water and air impacts, apply "beneficiary pays" principles to grid buildout, and weigh a Grid Acceleration Fund financed partly by developers, with possible requirements ranging from upfront grid contributions to dedicated battery storage. Gallup finds 70% of Americans oppose an AI data center in their area, 48% of them strongly - more local opposition than a nuclear power plant now draws. For anyone planning AI capacity, energy price, siting and community consent are first-order constraints, and roadmaps should assume friction.

The spicy one: the dashboard moat was authorization, and it just thinned

Flagged by the FinTech is Eating the World newsletter from Claude Code's documentation: AI-generated pages can now call MCP connectors at view time, with every call running under the viewer's own enterprise permissions. Two people open the same generated dashboard and see different Stripe transactions, different Jira queues - row-level security for throwaway apps. Viewers approve access before a page's first connector call, pages never see credentials, and enterprise admins can pre-authorize connectors through Okta. Gartner sizes the enterprise application spend exposed to agentic AI at up to $234 billion by 2030, roughly 20% of SaaS spending, and the vulnerable long tail is precisely the unglamorous stuff: reporting tools, admin consoles, BI layers, status pages. We rate this undervalued because it looks like a docs footnote while quietly rewiring who owns the interface to business data.

AI tip of the week

Route the task before you run the agent

Spend thirty seconds routing every task before you hand it to a coding agent. Quick mode for drafts and small diffs. High effort plus a mandatory human review for anything touching auth, payments, production data or anything painful to roll back. Decide in advance what evidence of "correct" looks like - a passing test, a diff you actually read, a screenshot - and give the agent a stop rule: if the evidence is thin, it halts instead of guessing. One addition after this week: never grant full system access by default. Every reported Codex deletion hit users who had switched sandboxing off.

The bottom line

The stack opened a little everywhere this week: models you can own, an operating system forced to host rival assistants, credentials agents can use but never read, dashboards that assemble themselves around permissions. Every one of those shifts rewards the same discipline - architecture built for choice. Pick components you can swap, govern the ones you keep, and the frontier arriving in open weights becomes an opportunity instead of a threat.

Sources

Newsletters: The Neuron · Exponential View · The Deep View · FinTech is Eating the World · The AI Break · What's Up in AI · AI Valley · The Batch

Web: Tom's Hardware on Kimi K3 · VentureBeat on Kimi K3 · Simon Willison on Kimi K3 · Thinking Machines: Introducing Inkling · Artificial Analysis on Inkling · TechCrunch on Inkling · European Commission DMA guidance · CNBC on the DMA decisions · Forbes on OpenAI's speaker · 1Password for Claude · Governor Hochul's Executive Order 62 · CNBC on the New York moratorium · Gallup on data center opposition · Gartner on agentic AI and SaaS spend · Claude Code artifacts documentation · Simon Willison on the Codex bug · The Decoder on Grok Build · MarkTechPost: K3 vs DeepSeek V4 Pro