Models
77 articles RSS
Microsoft Launches Seven In-House MAI Models, Built From Scratch Without Distillation to Cut OpenAI Reliance
Microsoft unveiled seven proprietary MAI models led by the 35B-active reasoning model MAI-Thinking-1, all trained from scratch without distillation as it reduces its dependence on OpenAI.
NVIDIA Open-Sources Nemotron 3 Ultra, a 550B Mamba-Transformer Mixture-of-Experts Built for Long-Running Agents
The 550-billion-parameter model activates 55 billion parameters per token, ships under the Linux Foundation's OpenMDW-1.1 license, and trades blows with China's Kimi-K2.6 on coding benchmarks.
Google Releases Gemma 4 12B, an Encoder-Free Multimodal Model With Native Audio That Runs on a 16GB Laptop
Google's new 12-billion-parameter open model drops separate vision and audio encoders, projecting raw image patches and audio waveforms straight into the LLM, and ships under Apache 2.0.
MiniMax Releases M3, an Open-Weight Model With a 1-Million-Token Context That It Says Tops GPT-5.5 on SWE-Bench Pro
Shanghai-based MiniMax launched M3 on June 1, pairing a 1-million-token context with a new sparse-attention design and company benchmarks that top GPT-5.5, with weights promised within 10 days.
Google Debuts Gemini Omni at I/O 2026, an Any-to-Any Model That Simulates the World to Generate Physics-Aware Video
Google DeepMind's Gemini Omni fuses Gemini reasoning with Veo, Genie, and Nano Banana to generate and conversationally edit video from any mix of text, image, audio, or video input.
Google Launches Gemini 3.5 Flash at I/O 2026, Beating Its Own Pro Model on Agentic and Coding Benchmarks
Google's new efficiency flagship outperforms Gemini 3.1 Pro on most evals while running 4x faster and costing 40% less.
Zyphra Releases ZAYA1-8B, an 8.4B-Parameter MoE Reasoning Model Trained End-to-End on 1,024 AMD MI300X GPUs
Zyphra's open-weight ZAYA1-8B uses 760M active parameters out of 8.4B total and was trained on a 1,024-GPU AMD Instinct MI300X cluster, narrowing the gap to frontier reasoning models on math benchmarks.
Miami Startup Subquadratic Emerges From Stealth With $29M and a 12-Million-Token Model It Says Beats Frontier Compute by 1,000x
Subquadratic launched SubQ on May 5, 2026 with a 12 million token context window and benchmarks it claims undercut Claude Opus by orders of magnitude. AI researchers are split between fascination and accusations of vaporware.
Mistral Medium 3.5 Folds Chat, Reasoning, and Coding Into a Single 128-Billion-Parameter Open-Weight Flagship
Mistral released Medium 3.5, a dense 128B open-weight model with a 256k context window that consolidates Medium 3.1, Magistral, and Devstral 2 under a Modified MIT license, with a per-query reasoning toggle.
Tencent Open-Sources HY-World 2.0, the First Foundation Model to Output Game-Engine-Ready 3D Worlds Instead of Video
Tencent's Hunyuan team released HY-World 2.0 on April 16, an open-source multi-modal foundation model that converts text or images into editable 3D assets importable into Unity, Unreal, and Isaac Sim — a sharp break from video-only world models like Google's Genie 3.
DeepSeek Releases V4 Under MIT License, Putting a 1.6-Trillion-Parameter Open Model Within Three to Six Months of the Frontier
DeepSeek's V4-Pro and V4-Flash arrive with 1M-token context, three reasoning modes, and pricing that undercuts frontier rivals by up to 8x.
OpenAI Releases GPT-5.5, the First Fully Retrained Base Model Since GPT-4.5, With 1M-Token Context and State-of-the-Art Agentic Benchmarks
OpenAI's GPT-5.5 arrives as a ground-up retrain with a 922K-token context window, 82.7% on Terminal-Bench 2.0, and two-tier pricing starting at $5/$30 per million tokens.