Large Language Models

ByteDance is reportedly training a new AI model with 10 trillion parameters, aiming to surpass Anthropic’s Mythos system and domestic Chinese competitors.

Moonshot AI has released Kimi K3, a groundbreaking 2.8T-parameter open model featuring native vision and a 1-million-token context window.

Security researchers have uncovered a new vulnerability called HalluSquatting that allows attackers to hijack AI coding assistants and build massive botnets.

MiniMax has teased its upcoming M3 model, featuring a custom sparse attention mechanism that delivers a 15.6x decoding speedup at a million tokens.

Google unveiled Gemini 3.5 Flash, a model designed for coding and agents that matches leading competitors’ performance at lower cost. The model is now available in Google Search and the Gemini App.

Cursor released Composer 2.5, an updated AI model offering 10x greater efficiency. The company doubles user access for one week and partners with SpaceXAI on a significantly larger model.

Google’s Gemini 3.1 Flash-Lite model is now generally available for enterprise use, designed for ultra-low latency and cost-efficient operations across software development, customer service, gaming, and financial applications.

xAI shipped Grok 4.3 out of beta with one million token context window, always-on reasoning, and Custom Voices voice cloning suite. The company also launched Imagine Agent Mode directly on the web interface.

Alibaba releases Qwen-Scope, an open suite of sparse autoencoders enabling direct manipulation of model features, data classification, and model behavior analysis without prompt engineering.

Meta Superintelligence Labs launched Muse Spark on April 9, 2026, its first large language model built to power Meta AI across apps, glasses, and the web. The model supports multimodal perception, parallel agents, and a new shopping mode starting in the US.

The US Department of Justice appealed a court ruling that blocked the Trump administration’s national security designation against AI company Anthropic. The dispute stems from a $200 million contract cancellation over military AI deployment rights.

Marc Andreessen joined the Latent Space podcast at a16z’s Sand Hill Road office to argue that AI represents decades of compounding progress, not a repeat hype cycle. He covered agents, open source, chip shortages, and why proving human identity online may soon require biometrics.

Alibaba’s Qwen3.6-Plus arrives with a 1M context window, top-tier agentic coding scores, and sharper multimodal vision. Open-source variants are coming soon.

Google has launched Gemma 4, its most capable open model family yet, available in four sizes from mobile-optimized edge models to a 31B dense model ranking third globally. Released under Apache 2.0, it supports 140+ languages and advanced agentic workflows.

Xiaomi released MiMo-V2-Pro, a 1-trillion parameter AI model, on March 18, 2026, ranking 10th globally on Artificial Analysis’s Intelligence Index at roughly one-seventh the cost of comparable Western models.

MiniMax has released its M2.7 model via the MiniMax Agent and API Platform, featuring reinforcement learning-based self-improvement and a 97% skill adherence rate across 40+ complex tasks.

Mistral AI has released Mistral Small 4, an open-source model combining reasoning, multimodal, and coding capabilities under the Apache 2.0 license. The model delivers a 40% reduction in completion time and three times more throughput than its predecessor.

AI token pricing has made “context” the central billing unit in enterprise AI, with model output costs ranging from $2.50 to $180 per million tokens across major providers. The market prices capacity but not the quality of relational structure within those tokens.

LinkedIn has deployed a new LLM-powered Feed ranking system serving 1.3 billion professionals, replacing a fragmented multi-source architecture with unified retrieval and a sequential Generative Recommender model.

NVIDIA released Nemotron 3 Super on March 11, 2026, a 120B-parameter hybrid MoE model with a 1 million-token context window designed for multi-agent AI applications. The fully open model scores 85.6% on PinchBench and delivers over 5x the throughput of its predecessor.

OpenAI released GPT-5.4 on February 12, 2026, introducing native computer-use capabilities and a 1-million-token context window. The model achieves an 83.0% win rate against industry professionals on the GDPval benchmark.

Google launched Gemini 3.1 Flash-Lite on March 3, 2026, offering developers a cost-efficient model priced at $0.25 per million input tokens with 2.5x faster response speeds than its predecessor.