Inkling opens weights, Grok Build opens code: July 16

Inkling gives technical teams a new AI model they can customize, Grok Build exposes the machinery inside an AI coding tool, and Reelful now assembles short videos from your own footage. In 30 minutes, you can choose one controlled test for each change and keep all three away from live customer work until they earn trust.

Inkling opens its model weights for custom AI jobs

Thinking Machines Lab released Inkling on July 15 with its full model weights available. Model weights are the learned settings that determine how an AI responds. The official Inkling announcement says it accepts text, images, and audio, while its Tinker service lets a technical team fine-tune it. Fine-tuning means teaching a general model a narrower behavior using your own examples.

This matters if a generic AI keeps mishandling one repeated, valuable job. Think product classification, brand-rule checks, or sorting recorded customer calls into fixed reasons. Inkling has 975 billion total parameters, the internal values learned during training, so running the full model yourself is specialist infrastructure work. A small business should test through a managed provider, not buy hardware around an unproven workflow.

The cost trap is customization without a scorecard. Write down what counts as correct before anyone trains the model. If you cannot grade an answer consistently, more training will produce a more confident mystery. The AI marketing tools stack helps you separate useful repeated jobs from experiments that do not deserve another system.

Grok Build publishes its coding-agent source

xAI published the Grok Build source on July 16 under the Apache 2.0 license. The official Grok Build repository contains the code for its terminal-based coding agent, a tool that reads a software project, edits files, runs commands, searches the web, and handles long jobs. The repository says it is synced periodically from xAI's internal code collection.

Opening the source gives your technical team a clearer route to inspect how the tool handles files, commands, saved checkpoints, and outside connections. It does not make the hosted AI model private, and it does not prove the tool is safe with your website or customer database. Owners with no custom software can ignore this. Businesses that let an AI edit a live website should act.

Create a copied test project with fake contact details. Ask the person responsible for your site to run one harmless change, then record every file changed and command executed. Keep automatic publishing turned off. Compare that log with the safeguards in your existing coding tool before replacing anything. For a broader selection test, use the AI automation tools comparison, where ownership and failure alerts matter more than a flashy demo.

Reelful turns camera-roll footage into finished short videos

Reelful's version 1.2.1 added a fully agent-style video workflow on July 14. Agent-style means the software plans and completes several steps from one instruction. The official App Store history says the iPhone app now plans, edits, captions, adds voiceover and music, then renders a finished social video from uploaded photos and clips.

The useful change is source control. You can start with real job footage, product shots, or an event recording instead of asking an image model to invent the business. Reelful lists monthly plans from $24.99 for about 10 videos, so the fair comparison is against editing time and revision waste, not against another writing subscription.

Build one 20-second test from five approved clips. In the prompt, name the audience, the order of the shots, the one claim you can prove, and the final action. Export it, then check every caption, spoken claim, logo, face, and music choice before posting. Do not upload customer footage without permission. If short video is already part of your week, compare this process with the AI avatar video workflow and keep the method that needs fewer correction passes.

On the bench

Inkling-Small is only a preview. Wait for full weights before estimating whether the lighter model belongs in a regular business workflow.

Grok Build's open code is worth a security review next. Check where prompts, files, and command logs travel before connecting a real repository.

Reelful needs a repeatability test. Run the same five clips through three prompts and count correction passes, not just finished videos.