Musk's Terafab Project to Speak at Internal ASML Event

Terafab, the SpaceX–Tesla joint venture led by Elon Musk, has been invited to an internal technical conference at ASML, the Dutch semiconductor equipment giant. ASML views Terafab as a "serious endeavor," and Musk is set to join remotely to present the project in detail.

Announced in March 2026, Terafab is a joint venture to manufacture cutting-edge chips for AI, robotics, and space data centers. The factory planned for the US is expected to cost **at least $55 billion** to build, a scale that could have a major impact on the semiconductor industry.

ASML's EUV (extreme ultraviolet) lithography technology is essential for cutting-edge semiconductor manufacturing, so Terafab deepening its ties with ASML can be seen as an important step toward making the project a reality. How far the massive $55 billion investment actually progresses, however, will depend on future fundraising and government support.

Police in England and Wales to Stop Using AI in Court Statements

Police forces in England and Wales have been instructed **to immediately stop using AI to draft evidentiary statements submitted in court**. The measure, reported by the UK's Financial Times, highlights growing concern over the reliability of AI in legal proceedings.

The decision reflects the risk that AI-generated text contains content that contradicts the facts (hallucinations). Court statements are treated as evidence, making accuracy critically important, and AI-generated output was judged not yet reliable enough to meet legal standards. The story also drew 43 points and 9 comments on Hacker News, fueling the debate over AI use in the justice system.

This move could become an important precedent in the broader effort by governments to balance AI adoption with legal and ethical safeguards.

Cohere Gives the Community Early Access to a Coding-Focused Model

**Cohere**, the Canadian AI company, has released its first coding-focused model as an early access drop for the LocalLLaMA community. Cohere co-founder Nick Frosst posted it directly on Reddit and is soliciting feedback from users.

The model, "**BLS-Mini-Code-1.0**," features:

Total parameters: 30B

Active parameters: 3B (believed to use a MoE architecture)

A lightweight design intended to run in local environments

Frosst says his team is "excited about the speed despite the small size," and the model reportedly achieves token output speed on par with others in its size class. The weights are publicly available on Hugging Face, and the plan is to move forward with community testing and improvements ahead of the official release.

Cohere has historically focused on large-scale enterprise models, and its entry into coding signals a strategic expansion. The 3B active parameter size appears aimed at competitiveness in the local inference market.

OpenCV 5.0 Released — DNN Engine Rewritten, Built-in LLM and VLM Support

**OpenCV**, the de facto standard image processing library, has released its major "5.0" update. The biggest change is that the DNN (deep neural network) engine has been completely rewritten, with **built-in support for LLMs (large language models) and VLMs (vision-language models)** added.

As a result, LLM/VLM inference can now be integrated into image processing pipelines with OpenCV alone. Shifting to a design where processing that previously required pairing with a separate inference framework (ONNX Runtime, TensorFlow, etc.) can be completed entirely within OpenCV could greatly simplify AI adoption on edge devices and embedded systems.

OpenCV is used extensively in industrial image processing around the world, and this update marks a turning point from an "image processing library" to an "AI vision integration platform."

Low-Bit GGUF Poised for Major Improvement — MoQ Quantization and the GSQ Format

New techniques, "MoQ (Mixed-precision Orthogonal Quantization)" and "GSQ (Group-wise Scaled Quantization)," show promise of fixing the **quality problems of low-bit (2-bit to 4-bit) quantized GGUF**, a long-standing challenge in the local LLM community.

Conventional low-bit GGUF dramatically reduces model size, but at the cost of noticeable degradation in output quality. MoQ automatically selects the optimal bit precision for each layer, while GSQ adjusts scaling on a per-group basis — mechanisms that maintain high quality even at low bit widths. The topic is gaining traction on Reddit's LocalLLaMA community, and if practical adoption progresses, large models could achieve high-quality inference even in environments with tight VRAM constraints.

---