Running a 27B Model on iPhone — PrismML's Compression Technology

PrismML, a startup backed by Khosla Ventures, has announced a breakthrough method for running a compressed version of Qwen-3.6-27B on the iPhone, drawing attention from the LocalLLaMA community on Reddit.

Being able to run a 27B-parameter model on a mobile device greatly expands the possibilities of local LLMs. Models runnable on mobile devices had long been considered limited to a few billion parameters, and PrismML's compression technique shows the potential to break through that barrier.

Detailed technical specifications and benchmark results have yet to be released, but the fact that a Khosla Ventures-backed startup is claiming the "largest AI model ever run on an iPhone" is worth noting.

Running Gemma 4 Inside Godot — A Challenge Using Only GDScript and Vulkan

A post on the LocalLLaMA community on Reddit has been making waves, describing how Google's Gemma 4 model was run directly inside the Godot game engine. The setup reportedly uses only GDScript (Godot's scripting language) and Vulkan compute shaders, with no dependence on external inference engines.

Running an LLM inside a game engine could change the paradigm of game development, enabling things like autonomous NPC dialogue and dynamic in-game content generation. Eliminating reliance on external APIs also makes it possible to achieve both privacy and offline operation.

The agent/ Folder Technique — Giving AI Agents "Project Memory"

An article on Qiita introduced the "agent/ folder" technique for retaining project context in AI agents such as Claude Code.

To address the problem of agents starting from a "first meeting" state every time a session opens, the technique consolidates the project's structure, decisions, and progress into an agent/ folder, allowing the agent to instantly reference past context. In measured cases, this reportedly reduced token consumption by as much as 30%.

As a practical approach to preventing CLAUDE.md bloat, this technique offers useful insights for developers who use AI coding tools daily.

Device Context Protocol — Bridging LLM Agents and Physical Devices

An open project called the "Device Context Protocol (DCP)" has been released on GitHub. It is a protocol for connecting LLM agents with physical devices, aiming to enable AI to directly operate IoT devices and other hardware.

While MCP (Model Context Protocol) handles connecting data sources and tools, DCP aims to standardize the physical device layer. Whether it can become a common standard for AI agents controlling real-world devices — the future direction of the ecosystem will be worth watching.

Anthropic Extends Free Access to Fable 5

According to The Decoder, Anthropic has extended free in-subscription access to Claude Fable 5 until July 19, 2026. The model was originally scheduled to move to pay-per-use, but the extension appears to be a response to intensifying price competition from OpenAI's GPT-5.6 Sol.

Subscribers can use Fable 5 up to 50% of their weekly usage cap. Price competition among AI models is a benefit for consumers, but it structurally puts pressure on providers' profitability.

---