Subject: Inkling drops, Grok goes open, and Reelful edits your camera roll A/B alternatives: - A) Apple's Qwen deal is live — what it means for your iPhone - B) 5 new AI tools you can actually use today

TL;DR: Open-weights momentum keeps building (Thinking Machines ships Inkling), xAI opens Grok Build under Apache 2.0 after a privacy scare, Apple Intelligence gets the green light in China via Alibaba's Qwen, and Reelful turns your camera roll into Reels with one prompt. Plus: Gemma 4 26B running on a 13-year-old Xeon at 5 tok/s.

The interesting pattern this week isn't a single model release — it's the layers underneath. A frontier-grade open-weights MoE. A coding-agent harness you can audit line by line. A CPU-only inference story that says the "you need a GPU" era is fading. And two AI products (Reelful, Rime) built on the assumption that voice and video are the new keyboard. Read the room.

1. Inkling by Thinking Machines · open-weights MoE built for fine-tuning

Thinking Machines (Mira Murati's lab) released Inkling, a 975B-parameter open-weights MoE (41B active, 1M context, multimodal) on July 15, 2026, available via their Tinker fine-tuning API. It's not the leaderboard king — it's pitched as "a good open-weights base for customization": controllable thinking effort, deep epistemics training (calibration + "I don't know" rewards), and agentic-coding features. ForecastBench-with-search score of 63.7 beats GPT-5.5's 64.7 within the margin and beats Claude Opus 4.8's 58.6.

Why it matters: Open-weights at frontier-grade quality, with the knobs turned the right way — if you want to fine-tune for your domain without a frontier-lab exclusivity contract, this is the most configurable foundation released in months.

Honest take: Inkling-Small (276B / 12B active) is the variant most teams can actually serve. The 975B is a research/hosted play, not something you'll self-host.

Source: https://thinkingmachines.ai/news/introducing-inkling/

2. Grok Build (open source) · xAI puts its CLI coding agent on Apache 2.0

xAI released xai-org/grok-build (the harness behind the grok CLI) under Apache 2.0 on July 15, 2026 — 844,530 lines of Rust, single commit, no history. The release came after users discovered the closed beta was silently uploading entire working directories to xAI's Google Cloud buckets. xAI disabled the default, deleted retained data, and dropped the source.

Why it matters: A full-screen TUI coding agent with interactive, headless, and ACP-editor modes, where every prompt, tool, and route is now readable. Useful for security reviews, custom agent forks, and for anyone burned enough to want a coding agent whose data path they can inspect.

Honest take: External contributions are explicitly not accepted — this is a transparency gesture, not a community project. Treat it as "auditable but read-only."

Source: https://github.com/xai-org/grok-build · coverage: https://simonwillison.net/2026/Jul/15/grok-build/

3. Apple Intelligence + Qwen · China's regulator greenlights Apple–Alibaba AI

The Cyberspace Administration of China approved Apple's AI services for launch in China on July 15, 2026, powered by Alibaba's Qwen models integrated into iOS, iPadOS, macOS, and visionOS. No rollout date yet. Apple had previously tried Baidu, DeepSeek, and ByteDance; Qwen is the survivor. Apple shares Q2 2026 Greater China revenue: $20.5B (up 28% YoY); Alibaba US shares jumped ~6% on the news.

Why it matters: Chinese iPhone users are about to get a generationally different Siri + system-wide AI than the rest of the world. If you build for iOS, plan for a forked Apple Intelligence surface in mainland China.

Honest take: The story is less about Qwen's quality and more about regulatory clearance — Apple's existing models (when shipped) will be censored/Qwen-routed. Don't expect feature parity across regions.

Source: https://techcrunch.com/2026/07/15/apple-intelligence-approved-for-launch-in-china-with-alibabas-qwen-ai/

4. Reelful · camera roll → TikTok-ready video from one prompt

Reelful (iOS, ex-Snapchat ML founder Kate Deyneka, in a16z's Speedrun) is an app that turns photos and clips from your camera roll into polished short-form videos. Pipeline: prompt → 30-second voice clone → pick media → AI plans the edit, writes the script, voices it, adds captions + music. You can also chat to refine (swap soundtrack, revise script). Pricing is credit or subscription ($25/mo Creator / $50/mo Pro / $100/mo Studio).

Why it matters: Aimed at founders, SMBs, and "I went to an event and want a Reel by the time I get home" workflows. Lands in a category where Opus Clip, Captions, and a handful of others already play.

Honest take: Convenient, but the brand risk is real — quality varies with source footage and prompt, and the AI-watermark is automatic. Don't ship it for a paying client without manual QA on the first ten outputs.

Source: https://techcrunch.com/2026/07/15/reelfuls-ai-turns-your-camera-roll-into-short-form-videos-for-social-media/

5. Gemma 4 26B on a 2013 Xeon at 5 tok/s · the "you don't need a GPU" proof point

A developer got Gemma 4 26B (MoE, Q8_0) running at ~5.2 tok/s decode / ~16 tok/s prompt eval on a $300 13-year-old HP StoreVirtual with dual Ivy Bridge Xeons, no GPU, AVX1 only (no AVX2/FMA3). The catch: had to patch ik_llama.cpp because two MoE graph ops fell through to default on instruction sets older than Haswell, producing "fluent-looking multilingual gibberish" with zero NaNs.

Why it matters: Concrete evidence that a useful local model can run on commodity, even retired, hardware — and a textbook story of where AI agents actually help (deep-diving a fork's dispatch logic and recognizing when the patch is right).

Honest take: 5 tok/s is reading speed, not building speed. Useful for local autocompletion, on-device summarization, or as a teaching artifact — not a Codex replacement. Treat the bug story as the real headline: an agent caught a silent uninitialized-tensor bug across ~240 tensors per forward pass.

Source: https://www.neomindlabs.com/2026/06/08/running-gemma-4-26b-at-5-tokens-sec-on-a-13-year-old-xeon-with-no-gpu/

Bonus link (not a full item): Rime raised a $24M Series A (M13 lead) for its enterprise voice-AI models — proprietary training data + phoneme-based architecture to handle brand names and noisy calls. Customers include Mayo Clinic, Dialpad, Upstart. Worth bookmarking if you ship voice agents. Source: https://techcrunch.com/2026/07/15/rime-picks-up-24m-series-a-to-help-enterprises-field-customer-calls/

Subscribe for weekly AI-tools recaps: https://buttondown.com/nexttool

Disclosure: Some links in this issue may be affiliate links, which means we may earn a small commission if you sign up. We only recommend tools we genuinely use or believe in.