Introduction

From the data collected on July 25, we've picked six stories, focusing on topics not yet covered in today's other digest articles. They're presented from a practitioner-oriented perspective: shrinking local inference footprints, real-world challenges of running agents, and how the new tokenizer affects Japanese.

---

Inflect v2: A 'Practically Usable' TTS at 4M Parameters

Developer Owen Song has released **Inflect v2**, an ultra-compact TTS that handles the entire text-to-speech pipeline in a single model. The lineup consists of two models: **Inflect-Nano-v2 (3.96M / 15.97MB FP32)** and **Inflect-Micro-v2 (9.36M / 37.53MB FP32)**. It outputs 24kHz audio without an external vocoder, and even on CPU, Nano achieves 10.72× real time while Micro reaches 6.28× real time.

In parameter count, that works out to roughly **1/21 of Kokoro and about 1/126 of Chatterbox**, and under 1/1,000 of Fish Audio S2 Pro. Naturally, this isn't a straightforward capability comparison — the point is to answer the question of how small you can go while remaining practical. That said, Micro scored 4.395 and Nano 4.386 on UTMOS22, and reportedly placed 2nd and 3rd in a community blind comparison. With fixed voices and English-only support, these models are now at a size where edge deployment is worth serious consideration.

Both Nano and Micro are available on Hugging Face

Code: github.com/owenawsong/Inflect

Interactive Playground: huggingface.co/spaces/owensong/Inflect-v2

Running Gemma 4 12B with 128K Context on Core ML

Okayuji on Zenn has released a **Core ML bundle for Apple Silicon Macs that handles Gemma 4 12B with up to 131,072 tokens**. The single 6.7GB bundle includes two functions — one for 32K and one for 128K — and the author measured **11 tok/s** with int4 decode on the GPU of an M4 Max.

Implementation highlights include a ring-buffer KV cache, copy-free mode promotion, and lossless speculative decoding built on Core ML, with every stage locked down by bit-exact verification. Since it runs directly on Hugging Face, it looks like a strong option for Mac users experimenting with long-document processing locally.

Testing Claude's New Tokenizer: Japanese Shrinks, English Grows +38%

Anthropic had indicated that the new tokenizer in Claude Opus 4.7 and later could consume **1× to 1.35× (up to a 35% increase) as many tokens** as the previous model. moname_ai on Zenn actually measured this 'depends on the content' range for both Japanese and English.

As the title suggests, the results show that **Japanese token counts trend downward, while English can reach +38%**. While the relative cost of Japanese processing drops, batch processing of English input and long prompts may force you to recalculate your API cost estimates. Before moving to production, it seems worthwhile to re-run measurements on your own primary workloads.

The `bash -c` Loophole in AI Agent Allowlists

Map on Zenn reports a structural hole lurking in allowlist management for commands executed by AI coding agents. Operators pile up per-command entries — 'auto-approve anything with this command name' — believing that's the safe side to err on. But if a generic executor that passes arbitrary strings straight to the shell (a so-called shell wrapper) gets mixed into the allowlist, **the per-command approvals effectively lose their meaning**.

As long as a path remains for passing arbitrary strings in a form like `bash -c "..."`, no matter how carefully you curate per-command allow/deny decisions, the net effect can end up equivalent to allowing everything. Before setting up Agentic CI, this is something to audit: allowlist granularity checked together with whether any path passes strings to a shell.

The AI Approval Pipeline That Burst at 300 Items a Month

JOINCLASS, who runs a one-person company with a fleet of AI agents, shares the experience of an approval pipeline set up as a safety mechanism that **broke down after exceeding 300 items per month**. Agents across 10 departments handled article writing, sales, accounting, and publishing, with every draft queued for approval before external actions. But human approval capacity couldn't keep up and it became the bottleneck. As a result, the system was reportedly redesigned into a two-tier structure: automated approval criteria with humans reviewing only exceptions.

This case makes it clear that a 98% automation rate isn't a victory of 'the system' — it stands on **design choices about where to insert human judgment**. It reflects a design shift underway, from 'delegate whatever AI can handle' to 'how do we dilute human approval'.

Open-Source LLM Rankings 2026: The 2.8-Trillion-Parameter Era

田口天晴 on Zenn has published a guide for choosing among open-weight LLMs as of July 2026. In the span of just two days, **Thinking Machines Lab**, founded by former OpenAI CTO Mira Murati, released its first model, **Inkling**. The next day, China's **Moonshot AI** announced **Kimi K3** with **2.8 trillion** total parameters, replacing what was then the world's largest open-weight model.

The article also touches on the point that 'size doesn't necessarily translate to usability,' offering a perspective that organizes comparison tables by practical use case (inference cost, hosting difficulty, license). It feels like we've entered the moment to shift from 'the biggest, by default' to 'what fits the purpose'.

---

Wrapping Up

Today's lineup centered on topics with direct operational impact: shrinking local inference footprints, structural challenges in agent operations, and how the tokenizer change affects Japanese costs. We'll keep you posted as further important updates arrive tomorrow.