1. Thinking Machines Lab Releases Inkling, Its First Public Multimodal Model
Thinking Machines Lab, which previously previewed its multimodal interaction architecture in May, has launched its first public model, Inkling. This release marks the transition from research preview to a production-ready, open-weights model featuring 975 billion total parameters and controllable reasoning effort. Trained from scratch on 45 trillion tokens of multimodal data, the model utilizes a 66-layer decoder-only transformer with 256 routed experts and 2 shared experts.
- • Inkling is an open-weights, natively multimodal Mixture-of-Experts (MoE) model featuring 975 billion total parameters and 41 billion active parameters.
- • The model features a controllable thinking effort mechanism, allowing developers to programmatically adjust the reasoning budget between 0.2 and 0.99.
- • It supports a 1-million-token context window and accepts text, image, and audio inputs.
- • The model is released under an Apache 2.0 license, with weights available on Hugging Face.
- • Inkling is accessible via hosted APIs from partners including TogetherAI, Fireworks, Modal, Databricks, and Baseten.
This release makes the company's multimodal technology available for public use, providing developers with a high-parameter model that builds on the interaction research previously announced.
2. Soofi Consortium Releases Soofi S 30B-A3B Hybrid MoE Model
The Soofi Consortium, coordinated by KI Bundesverband, has released the preview weights for Soofi S 30B-A3B on Hugging Face. Built as a hybrid Mamba-Transformer MoE, the model combines Mamba-2 sequence-mixing layers with granular MoE and Grouped-Query Attention layers to optimize performance and efficiency. The training process utilized a Warmup–Stable–Decay schedule across 26.68 trillion tokens, with the German language share of training data increasing to 15.32% in the second phase.
- • Soofi S 30B-A3B is an open-weights hybrid Mamba-Transformer Mixture-of-Experts (MoE) model with 31.6 billion total parameters and 3.2 billion active parameters per token.
- • The model supports a context window of up to 1 million tokens and is available for deployment via vLLM.
- • The architecture features 52 layers, including 23 Mamba-2 sequence-mixing layers, 23 granular MoE layers, and 6 Grouped-Query Attention layers.
- • The model achieved an English aggregate score of 70.1 and a German aggregate score of 79.1 in evaluations against other open base models.
- • Training was funded by the German Federal Ministry for Economic Affairs and Energy and conducted on Deutsche Telekom's Industrial AI Cloud using up to 512 NVIDIA B200 GPUs.
It allows developers to deploy a highly efficient, open-weights German/English hybrid model with a 1-million-token context window using vLLM.
3. Google Officially Releases Gemma 4 E2B Optimized for Pixel 10 TPU
Building on the release of the Gemma 4 12B model and subsequent community efforts to run it on mobile hardware, Google has now introduced Gemma 4 E2B. This version is specifically optimized for the Pixel 10's Tensor Processing Unit (TPU), enabling native, low-latency, and offline multimodal capabilities that go beyond the previous community-led implementations.
- • Gemma 4 E2B is an official, TPU-optimized version of the Gemma 4 architecture.
- • Enables native, offline multimodal tasks including audio transcription and image identification.
- • Supports direct control of system-level phone functions via private voice or text commands.
- • Transitions from experimental community-led local inference to a supported, hardware-accelerated native feature.
This official optimization provides a stable, high-performance path for developers to integrate private, on-device multimodal AI into Pixel 10 applications, moving from experimental community setups to a supported native feature.
4. SpaceXAI Open-Sources Grok Build Terminal Coding Agent
SpaceXAI has released the source code for Grok Build, transitioning the terminal-based developer tool from its previous status as a paid, closed-source beta to an open-source project. Operating as a full-screen text user interface (TUI), Grok Build can be run interactively or headlessly within CI pipelines. It integrates with external editors using the standardized Agent Client Protocol, making it a flexible addition to modern development environments.
- • Grok Build is now open-sourced under the Apache License, Version 2.0.
- • The tool functions as a full-screen TUI for terminal-based AI coding.
- • It supports interactive use, headless operation for CI, and editor integration via the Agent Client Protocol.
- • Prebuilt binaries are available for macOS, Linux, and Windows, with source builds requiring Rust and protoc.
- • External contributions are not accepted despite the open-source license.
The shift to open-source allows developers to access and modify the terminal-based AI coding agent, which previously required a paid subscription to use.
5. deja Indexes Local Coding Agent History for SSH-Synced Memory
Unlike memory platforms that capture new data in real-time, deja indexes existing agent logs, including historical sessions created before its installation. It provides a secure, local memory layer that integrates with Model Context Protocol (MCP) workflows. The tool can be installed via Go, npm, Homebrew, or a shell script.
- • deja is a zero-dependency Go binary that indexes local conversation histories from Claude Code, Codex, and opencode.
- • The tool performs incremental indexing locally, storing data in an inverted index at ~/.cache/deja.
- • It automatically redacts sensitive credentials like API keys, tokens, and JWTs during indexing.
- • Features include MCP recall, session-start auto-recall, and synchronization between machines via shared folders or SSH.
- • deja operates entirely locally with no network code, ensuring data privacy, and is licensed under the MIT license.
Gives developers a local, secure, and searchable memory layer for coding agents like Claude Code that automatically syncs across machines and redacts sensitive credentials.
6. Google Releases LiteRT.js for WebGPU-Accelerated Browser Inference
Google's release of LiteRT.js brings high-performance, hardware-accelerated on-device inference to web browsers. The library requires full model delegation to an accelerator, falling back to CPU execution if a model cannot be fully delegated. Google recommends using LiteRT.js for model inference while continuing to use TensorFlow.js for pre- and post-processing tasks.
- • LiteRT.js compiles the native LiteRT (formerly TensorFlow Lite) runtime to WebAssembly to execute models directly in the browser.
- • The library utilizes three backends: XNNPACK for CPU, ML Drift via WebGPU for GPU, and the experimental WebNN API for NPUs.
- • Google reports up to 3x faster performance than other web runtimes, and a 5–60x speedup on GPU/NPU backends compared to CPU.
- • PyTorch models can be converted to .tflite format using LiteRT Torch if they are TorchDynamo-exportable and have static dimensions.
- • Developers must manually manage tensor memory by calling delete to avoid memory leaks in LiteRT.js.
It enables web developers to run hardware-accelerated on-device AI models directly in the browser with up to 3x faster performance using WebGPU and WebAssembly.
7. ExLlamaV3 Reaches Production v1.0.0 with Major Performance Upgrades
Following over a year of development, ExLlamaV3 has reached its first production-ready release. The update focuses on performance and ease of deployment, eliminating the need for external flash-attention-2 and xformers libraries. It also introduces a Mixture-of-Experts kernel ticket scheduler and support for GptOssForCausalLM and NemotronHForCausalLM models, alongside faster compilation times.
- • ExLlamaV3 v1.0.0 removes dependencies on flash-attention-2 and xformers, simplifying the local build and deployment process.
- • Tensor-parallel support has been extended to most models, including Gemma 4.
- • A new attention kernel supports online cache quantization and dual input for SWA layers and attention sinks.
- • The release introduces a new conv1d kernel, eliminating the need for causal_conv1d.
- • Performance improvements include enhanced GEMM/GEMV performance on Ampere architecture and a new INT8 GEMV kernel.
It accelerates local LLM inference and simplifies deployment by removing heavy dependencies like flash-attention-2 while adding tensor-parallel support for models like Gemma 4.
8. Anthropic Mitigates Claude Memory Exfiltration Vulnerability
A security researcher demonstrated a data exfiltration attack against Claude's memory system, which stores high-fidelity user profiles via a daily summarization pass and retrieval tools. By tricking Claude into visiting a malicious website, the attacker could force the web_fetch tool to follow a directory structure of links, encoding and leaking sensitive user data. Anthropic has mitigated the threat by restricting web_fetch from following external links.
- • Researcher Ayush Paul discovered a vulnerability in Claude's memory system that allowed exfiltration of stored user profiles.
- • The attack exploited Claude's web_fetch tool, forcing the model to encode and transmit sensitive data through URL paths on malicious sites.
- • The vulnerability could be triggered if a user asked Claude to visit a malicious site or if the site appeared in search results.
- • Anthropic mitigated the issue by disabling web_fetch's ability to follow links on external pages, limiting navigation to search results and user-provided URLs.
- • The vulnerability was disclosed via Anthropic's HackerOne bug bounty program, though no bounty was awarded as the issue had been identified internally.
It informs developers building web-browsing agents about critical data exfiltration vectors and explains Anthropic's mitigation of disabling external link-following in Claude's web_fetch tool.
9. Ambiance LLM Harness Uses Unix Filesystem as Agent Interface
Ambiance reimagines the agent runtime environment by treating the local filesystem as the primary interface for storing data, logs, and configurations. By using an event-driven kernel to monitor filesystem changes, the harness ensures that agents remain responsive to external state changes, offering a structured alternative to traditional agent loops.
- • Ambiance leverages an agent's existing knowledge of coding and systems administration to reduce cognitive load.
- • The system features a 'Kernel' that acts as an event bus, monitoring the filesystem for changes to trigger LLM responses.
- • It defines three default LLM instances: 'root' for system tasks, 'pai' for human interaction, and 'librarian' for journaling.
- • The project was initiated at the Recurse Center and is available for testing via whitematterlabs.ai or an installation script.
It provides an experimental, Unix-inspired local environment and event-driven kernel for building more reliable and responsive LLM agents.