Company@X — 2026-06-07#

Signal of the Day#

The llama.cpp inference framework has officially merged support for Gemma 4’s Multi-Token Prediction (MTP). This integration represents a major step forward for local AI deployments, allowing developers to combine Quantization-Aware Training (QAT) with MTP to run complex models at the edge with significantly faster processing speeds and lower hardware requirements.

Key Announcements#

[llama.cpp / Hugging Face] · Source Gemma 4’s MTP architecture has been officially integrated into the popular llama.cpp repository. This allows developers and researchers to leverage Gemma 4 QAT alongside MTP, delivering a lightweight, highly optimized setup for local execution. The development highlights the industry’s continued momentum toward high-performance, low-latency models that bypass the need for cloud compute.

[Nvidia] · Source Nvidia continues to aggressively position itself as a dominant force in the open-source ecosystem, currently publishing 9 of the top 30 trending models on the Hugging Face front page. Furthermore, Nvidia’s research division released a new leading paper titled “Cosmos 3: Omnimodal World Models for Physical AI,” signaling deepened strategic investments in multi-modal architectures for embodied robotics.

[Tesla] · Source Tesla amplified the operational capabilities of its Full Self-Driving (Supervised) software, emphasizing its ability to autonomously execute lane changes, pass traffic, and self-park in complex scenarios with “eerie confidence”. The company is also heavily promoting its off-grid lifestyle utility, highlighting that its standard “Camp Mode” allows vehicles to sustain climate control, device charging, and media playback over multiple uninterrupted days.

[Two Labs AI] · Source Y Combinator-backed robotics startup Two Labs AI unveiled an initial look at their hardware in action with “Episode 1: Sardor’s Birthday”. The launch video focuses on demonstrating the company’s core objective: engineering robots capable of naturally integrating into human environments and interacting seamlessly with people.

Also Noted#

• [Hugging Face] (Source): The platform is currently running a “Build Small Hackathon” to incentivize the creation of localized, practical AI applications for non-technical users.
• [Community AI] (Source): Open-source developers launched “Super Gemma 4 26B Uncensored GGUF v2,” an optimized local model boasting 90% faster prompt processing that runs on consumer hardware with 16 to 22 GB of VRAM.
• [Pollen Robotics] (Source): Developers successfully deployed the Reachy mini robot running locally in near real-time using a lightweight stack that includes Gemma 4 E4B QAT and Qwen3-TTS models.
• [Microsoft] (Source): Microsoft signaled its continued technical partnership and integration with the Mercedes-AMG F1 team following their performance at the Monaco Grand Prix.
• [Y Combinator] (Source): VCs are signaling that the next major frontier beyond scaling model size is the development of coding agents that can write executable world models to achieve maximum skill-acquisition efficiency.