Models & Releases
Meta released Muse Spark Open, now the strongest American-made open-weight model, significantly closing the performance gap with Chinese frontier models.
DeepSeek released V4.1 Flash 1017, a sparse MoE model with CED architecture offering cost-efficient performance for coding agents and long-context tasks.
Microsoft has pre-trained MAI-Base-2, a new MoE foundation model with 2.8 trillion total parameters that achieves 4-12x better pre-training efficiency than DeepSeek-V4-Pro-Base.
Meta releases Muse Spark Open, an 816B-parameter multimodal MoE model with open weights that scores highest among US open-weights models on Artificial Analysis' Intelligence Index.
Reflection releases Beam, a 501B parameter open-source MoE model optimized for coding and agentic use cases with 1M token context and Apache 2.0 license.
Elorian released Glimpse, an AI model that reasons with images by repeatedly zooming and cropping during analysis rather than converting to tokens once.
Nvidia released Nemotron 3.5 models that run on-device while supporting open source, reinforcing their GPU infrastructure strategy.
Perplexity AI has released pplx-decider-flash-v1.2 model weights on Hugging Face, supporting multimodal inputs and tool calling functionality.
IBM releases Granite Switch 4.1 models with integrated chat, decision-making, and speech capabilities that share GPU memory via cache-efficient aLoRA adapters.
Hugging Face announced a new 501B parameter open-weight model with only 23B active parameters per token during a live stream event.
FLUX.3 Image Klein, Black Forest Labs' latest open image generation model, is now available in Diffusers with layout control via bounding box prompts and multi-reference image-to-image capabilities.
Apertus 1.5 adds multimodal image/audio understanding, 262k context window, thinking mode, and improved tool use to the open-source 8B/70B models.
Google DeepMind announces Gemini 4 Argon is accelerating breakthroughs in quantum computing, nuclear fusion, and protein design with specific technical achievements.
Reflection's Beam, Google's EmbeddingGemma 2, and Mistral Large 4 represent three complementary approaches to building more capable AI systems through efficient reasoning, multimodal search, and visual understanding.
Dyna Robotics has released Dyna-2.1, a physical agent that achieves hour-long autonomy by combining a world-action model trained on 1M hours of human video with custom semi-humanoid hardware.
Higgsfield extends AI video generation from 1 minute to over 10 minutes using GPT-6 Astra, GPT-Image-2.5, and GPT-Live technologies.
LiquidAI releases LFM2.5-24B-A2B, a 24B-parameter reasoning model with only 2B active parameters for efficient on-device agent execution.
Google's Embedding Gemma 2 model is trending on Hugging Face, offering efficient embeddings for multimodal search and retrieval applications.
Google released Gemini 4 Argon globally for coding and enterprise workflows, with API pricing starting at $2/$10 per million tokens and benchmarks showing 52.8% on CursorBench.
Meta releases Muse Spark Open, an open-weight 816B parameter multimodal MoE model optimized for agentic coding with text, image, audio, and video capabilities.
Meta and Apple collaborated to optimize Muse Spark Open AI model for Mac Studios using MLX framework, enabling desktop deployment.
Sam Altman teases a milestone shirt featuring chain-of-thought reasoning from an AI solution to Navier-Stokes equations.
Meta launches Muse Spark Open, an 816B parameter open-weight AI model designed to run natively on Macs for coding and agentic workflows.
NVIDIA released Nemotron 3.5 Super VL (120B parameters) and two Bolt edge models optimized for local agentic workloads across data center to desktop deployments.
Meta has released Muse-Spark-Open-816B-A42B, a massive 816 billion parameter model with advanced reasoning capabilities and multimodal support.
Mistral AI's 'Le Chonk' model achieves 42% on Dense 200 visual grounding benchmark, outperforming even frontier closed models.
Google DeepMind announced Gemini 4 Argon, a frontier AI model that accelerates robotics, cybersecurity, and software engineering tasks, with benchmarks showing top performance on CWE-Bench.
Hugging Face CEO teases major model release during 'fashion week'; community speculates on Qwen 4.0 or similar heavyweight competitor.
Microsoft launches Maia 300, a custom AI inference chip designed for Azure deployments, scaling to multi-gigawatt capacity.
Google's Gemini 4 Argon model is now available via API, priced at $2/M input tokens and $10/M output tokens, with claimed performance improvements across coding, finance, and legal domains.