AI News of the Week (2nd January)
TLDR: Key AI Developments This Week
Deepseek unveiled a novel training method for more efficient LLM development, Z.ai open-sourced GLM-4.7 for agentic workflows, and SK Telecom launched Korea's first sovereign 500B-class enterprise LLM, while OpenAI announced a shift toward audio-first hardware. Meanwhile, Meta acquired digital employee startup Manus, NVIDIA advanced AI21 Labs acquisition talks and licensed Groq technology for inference acceleration, and OpenAI declared 2025 as the year of agent-native APIs.
Deepseek Unveils "Manifold-Constrained Hyper-Connections" (mHC) Training Method
Chinese AI lab Deepseek has introduced a novel training technique called Manifold-Constrained Hyper-Connections (mHC), designed to significantly reduce the cost and improve the efficiency of training large language models. Building on earlier "Hyper-Connection" concepts, this method reportedly enhances stability and scalability without increasing computational overhead. The technology has already been tested on models with up to 27 billion parameters, with industry experts suggesting it could pave the way for Deepseek's next major model release. Read more
OpenAI Reportedly shifting focus to "Audio-First" Hardware for 2026
Reports emerging in early January suggest OpenAI is intensifying its development of a new consumer device that prioritizes voice interaction over traditional screens. Collaborating with former Apple design chief Jony Ive, the company is reportedly aiming to launch an audio-centric device in 2026. This strategic shift is supported by internal restructuring, bringing together engineering, product, and research teams to build more natural-sounding audio models that surpass current capabilities in speed and expressiveness. Read more
Z.ai Open-Sources GLM-4.7 for Agentic, Tool-Using Development
Z.ai has released GLM-4.7, an open-source foundation model designed specifically for real-world, multi-step development workflows where tool use and iterative reasoning matter. Unlike chat-oriented models, GLM-4.7 emphasizes consistent “think-then-act” behavior for agent-style coding environments, making it well suited for long-running development tasks. With open weights available via Hugging Face, the release gives teams a self-hostable, auditable alternative to closed coding models, especially appealing for organizations prioritizing control, transparency, and on-prem deployment. Read more
SK Telecom Launches A.X K1, a Sovereign 500B-Class Enterprise LLM
SK Telecom has unveiled A.X K1 and A.X K1 Mini, a 500B-class large language model family optimized for Korean-first enterprise use. The models emphasize deep alignment with Korean language, culture, and regulatory requirements, offering an alternative to globally generic cloud APIs. For enterprises that need frontier-scale reasoning while retaining tighter control over localization, compliance, and deployment, A.X K1 signals the growing importance of sovereign and region-specific AI infrastructure. Read more
Meta Expands Into “Digital Employee” Agents With Manus Acquisition
Meta is moving deeper into agent automation by acquiring and commercializing Manus, a general-purpose AI “digital employee” designed for research and operational tasks with minimal prompting. Rather than positioning agents as standalone tools, Meta plans to integrate Manus directly across its consumer and enterprise products, highlighting a broader platform strategy where agent labor becomes a native capability of large ecosystems. For builders, this reinforces the shift toward bundled, platform-owned agents rather than independent automation layers. Read more
NVIDIA’s AI21 Talks Signal Vertical Consolidation of the AI Stack
NVIDIA is reportedly in advanced talks to acquire AI21 Labs, signaling a push beyond hardware into foundation-model IP and enterprise AI software. If completed, the deal would further compress the AI stack, combining compute, inference optimization, and first-party models under one roof. For developers and enterprises, this trend suggests future AI platforms may increasingly arrive as vertically integrated ecosystems rather than modular, mix-and-match components. Read more
NVIDIA Licenses Groq Tech to Accelerate Inference Performance
NVIDIA has licensed inference technology from Groq and hired key Groq leaders, underscoring how critical inference speed and cost efficiency have become in large-scale AI deployments. Instead of relying solely on internal development, NVIDIA is buying both IP and talent to compress its inference roadmap. For teams running agentic workloads at scale, this points to rapid improvements ahead in latency, throughput, and serving cost, especially as inference becomes the dominant operational expense.
OpenAI Frames 2025 as the Year of Agent-Native APIs
OpenAI’s year-end developer roundup positions 2025 as the transition from prompt-driven interactions to production-grade AI agents. The post highlights standardized primitives such as the Responses API, Agents SDK, tool calling, and Codex workflows, reducing the custom orchestration glue teams previously had to build themselves. For developers planning 2026 roadmaps, the message is clear: durable agent threads, evaluation loops, and tool-first design are becoming the default architecture for serious AI systems. Read more
NVIDIA’s Startup Investments Reveal the Future AI Stack
A TechCrunch analysis of NVIDIA-backed startups shows how the company is shaping the AI ecosystem end-to-end, from chips and inference engines to agent platforms and application layers. These investments provide an early signal of where NVIDIA expects long-term value to concentrate. For founders and engineering leaders, this matters because platform dependencies, partnership opportunities, and even acquisition paths are increasingly influenced by NVIDIA’s strategic view of the full AI stack. Read more
