AI Model Release Tracker

• Unified design: 120B MoE model (12B active) handles text, images, documents, video, tool calls and coding in a single checkpoint
• Long context &...

• Open access push: Released open-weight Muse Spark (816B params) for coding and agents, runnable on two M5 Ultras and hosted on Hugging Face
  -...

Krea 3 stands out for preserving product identity in commercial image workflows.

• Keeps labels, logos, and packaging exactly as-is while generating...

• Claude Opus 5.5 posts multiple supported scores while MAI-Transcribe-2 shows zero public rankings, leaving every category incomparable.
• Claude...

NVIDIA's new 120B-parameter (12B active) multimodal model brings strong image and video understanding to open weights.

• HF availability lets users...

FLUX 3 Image Klein combines 5.6B parameters, open weights, native 4K output, pixel-perfect localized editing via bounding boxes, and control over up...

Ant Group open-sourced Ling-3.1-flash weights under the permissive MIT license, extending the lineage after July's Ling-3.0-flash and September's multimodal Ling-3.0-flash-VL.

Liquid AI released LFM2.5-24B-A2B on October 13, 2026 — a 24-billion-parameter MoE model that activates only 2.45 billion parameters per token, targeting on-device agent deployment with low inference costs.

LensErase introduces a mask-based foundation model pretrained from the Lens text-to-image model for broad real-world erasure tasks.

• Scale: Built on...

Google's new Gemini 4 Argon claims an industry-leading 1 million token output limit—roughly 8x rivals—targeting long-context agent workflows and code...

• Drama 3 is a 35B MoE text-to-speech model with ~3B active parameters per token, released with open weights on Hugging Face.
• It accepts...

VideoJev replaces the language-model head with a candidate scorer that outputs intervals or option scores directly, skipping autoregressive...

Neuron AI fuses 3D vision, touch, audio, robot state, language and memory through a Perception-Language-Action architecture that aligns all inputs...

• TacitVLA keeps reasoning supervision during training to improve backbone representations but removes current-step reasoning text from the action...

DEMON provides direct Hugging Face paths to Stable Audio 3 checkpoints and a unified real-time demo.

• Deployment: huggingface-cli download pulls...

Tether released QVAC TranscribePsy, a 254M-parameter open-source multilingual ASR model that runs entirely on-device for smartphones, smart glasses,...

Liquid AI’s 24B MoE with 2B-active params targets laptop and PC agents through harness-trained tool use and fast on-device inference.

• 2B-active...

• Sparse efficiency: 35B total params with only ~3B active per token via MoE and FP8 experts.
• Hardware barriers: Needs 48 GB NVIDIA FP8 GPU; CPU,...

ViewCraft jointly evolves an executable evidence interface and reading procedure, improving multimodal-agent task scores by an average of 6.1 points...

Meta has released its first frontier AI models and is re-entering the frontier race.