Inkling 2 Small goes open-weight, Anthropic lets you self-host Claude Managed Agents, and Mirendil raises $1B

Nov 12, 2026

OPEN WEIGHTS: Thinking Machines puts Inkling 2 Small on Hugging Face under Apache 2.0, and the lab's own table puts it first on game-dev and 3D design arenas

Thinking Machines Lab updated its Inkling 2 Small announcement on Nov 12 to say the model's weights are now downloadable on Hugging Face, a week after it launched the model with closed weights on its Tinker platform while outside safety testing and red-teaming finished. It is a mixture-of-experts model with 594B total parameters and 17B active, with image input and adjustable effort. The lab says it was trained with reinforcement learning on hundreds of thousands of long-horizon coding tasks across several agent harnesses, and that it checks its own work without being asked. Demos include a procedurally generated Three.js city-builder game and a Blender scene. In the lab's own comparison table, it posts the best scores among the listed models on DesignArena's Game Dev (1424) and 3D Design (1457) boards, but it trails DeepSeek-V4.1-Flash on several coding benchmarks, and the lab says it "doesn't always choose the best tools for the job." Day-one support covers vLLM, SGLang, Unsloth and TokenSpeed, with hosted APIs at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. Running it yourself is not a laptop job: the model card calls for at least 500 GB of GPU memory for the recommended NVFP4 version. Most benchmark numbers are compiled by the lab, so test it on your own repo.

Source: https://thinkingmachines.ai/news/inkling-2-small/

Nov 12, 2026

AGENTS: Anthropic lets companies self-host the whole Claude Managed Agents stack, from harness to session history

Anthropic opened a private beta of self-hosted Claude Managed Agents on Nov 12. Until now, teams could move only the sandboxes where agent tools run onto their own machines. Now the whole service can run in a company's own cloud or data center: the agent harness, the server and its database, the workers that run shell and file tools, credentials and session history. Model calls still go out to Claude, either directly to Anthropic, through Amazon Bedrock, Google's Agent Platform or Microsoft Foundry, or through the company's own gateway, so this is self-hosted orchestration, not an on-premises model. Anthropic pitches it for agents that run unattended on internal code, such as fixing broken builds, testing releases and working through bug backlogs. The design keeps the agent loop and secrets outside the sandbox, can lock each environment to a list of allowed hosts, and logs every tool call over OpenTelemetry. Access is through Anthropic account teams for eligible organizations only. AI Frontier Post reports that scheduled runs and webhooks are missing from the beta, nothing migrates from the hosted service, and no pricing has been published. This lands one day after Google added bring-your-own sandboxes to its own managed agents.

Source: https://claude.com/resources/articles/self-hosted-claude-managed-agents

Nov 12, 2026

CODING TOOLS: Expo Simulators gives coding agents cloud iPhone and Android devices, so they can prove a mobile fix works

Expo launched Expo Simulators on all paid plans on Nov 12. Your coding agent, CI job or teammate can start a cloud iOS Simulator or Android Emulator with one command, with no Mac needed, so you can run the iOS simulator from Linux, from Windows through WSL, or from inside a cloud sandbox. Each session runs on its own virtual machine, so parallel agents don't clash, and it records a video of the whole run plus screenshots, logs, network traffic and CPU and memory use. The intended loop, per Expo's release: a bug report goes to an agent, which reproduces it, fixes it, tests the fix on cloud devices and attaches the recording to the pull request. It works with Cursor, Claude Code and Codex, and adds no AI model of its own. Apps don't have to be built with Expo: Expo lists React Native, Swift, Kotlin, Flutter and Unity builds as supported, which matters for mobile game devs. Install the skill with npx skills@latest add expo/skills --skill eas-simulator. Pricing is $0.10 per simulator minute, and the Starter, Production and Enterprise plans include 60 minutes a month until 2027.

Source: https://expo.dev/blog/introducing-expo-simulators-cloud-simulators-built-for-agents

Nov 12, 2026

CODING TOOLS: Vercel and shadcn release UI SDK, 50+ interface behaviors your agent assembles into components for any framework

Vercel introduced UI SDK on Nov 12, built by shadcn (of shadcn/ui) and Shu Ding. The idea, as shadcn explains it: instead of handing an agent a finished Select component, give it the pieces a Select is made of, such as the popup, the selection logic and type-to-jump. Then it can build a Select, but also an autocomplete, a combobox or a component nobody has designed yet. The UI SDK site lists more than 50 behaviors, including keyboard focus, popups, multi-select, drag-to-reorder, resizing, pan-and-zoom canvases and virtualized lists, plus a starter component library. Everything works in React, Vue, Svelte, Solid and Web Components. You add the UI SDK skill to your agent and ask it for what you need, and the behaviors bring keyboard navigation and screen reader support that agent-written UIs often skip. Vercel also says agents can operate the finished controls by reading the page rather than taking screenshots, which it claims is more than twice as fast at under half the cost (company figure). The code is on GitHub under Apache 2.0 and is described as a public beta. For game devs building in-browser editors and tools, the timeline, canvas and resize behaviors are worth a look.

Source: https://ui-sdk.dev

Nov 12, 2026

GAME DEV: World Labs opens its Atlas world model to everyone, with exports to splats and meshes for Blender and game engines

World Labs said on Nov 12 that it is "opening up access to Atlas to everyone today," with apps and API endpoints that turn an idea for a 3D world into keyframes, video clips or 3D Gaussian splats. Atlas, first shown in September to a small group of partners, now has two browser apps. Viewfinder builds a scene from one image, or reconstructs a real place from up to 64 photos. Chisel starts from simple blocks and planes you lay out as walls and furniture, then styles them from a text prompt or reference image. According to AlphaSignal's review of the API docs, endpoints include scene generation, camera pose and depth estimation, and a splats2Mesh converter that turns a splat into a polygon mesh you can bring into Blender or Unreal Engine. The API is still labeled beta, there is no dedicated pricing page yet, and the company's quality claims have not been independently reproduced. For indie teams, the practical use today is fast greyboxing and reference scenes: block out a level in Chisel, export it, and rebuild it properly in your engine. World Labs announced in September that it is joining AMD.

Source: https://x.com/theworldlabs/status/2120942033401717108

Nov 12, 2026

DEALS: Mirendil raises $1B to give small teams frontier AI labs run by agents, and Agile Robots raises $500M at $8.5B

Mirendil announced a $1B Series A on Nov 12, led by Kleiner Perkins and Clearlake Capital, with Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also in. The San Francisco lab, founded in June by researchers from Anthropic, Google DeepMind, OpenAI and xAI, wants one person with agents to do work that used to take a whole research team. Its proof so far: one researcher working with agents pretrained Mirendil-Base, a competitive coding model, from scratch in 3.5 months on under $10M of compute. Mirendil then post-trained it into Fern, which holds up to 10M tokens of context, enough for days of agent work. Its Lor platform, now in private beta, runs hundreds of agents writing code and running experiments for days at a time (all company claims). Dealroom notes that no valuation was disclosed, and that the round comes about five months after a $200M seed. In robotics, Munich's Agile Robots closed a $500M Series F at an $8.5B valuation, led by SoftBank with NVIDIA, Hitachi and T.Capital, which brings its total funding to $2B. The German Aerospace Center spin-off says it has installed more than 20,000 robotic solutions, and its plant near Munich can build up to 10,000 humanoid robots a year. The new money goes to U.S. expansion and more production in Germany.

Source: https://mirendil.com/news/series-a/

Ship faster this week

Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.