AI Model Release Tracker

Flagship Open Model Releases

• 🔥 Kimi K3: Moonshot AI released Kimi K3, a 2.8T-parameter open-weights multimodal model with 1M context that leads...

Moonshot AI released Kimi K3, a 2.8-trillion-parameter open model with 1M context and native multimodal support, featuring architectural advances for...

Gemini 3.5 Flash launches as a high-speed multimodal model optimized for agentic workflows and complex coding.

• 1M-token context processes hour-long...

WanSong v1.0 delivers a pure diffusion-based model that generates high-fidelity, multilingual songs up to 5 minutes while outputting dual stems in a...

MLLMs are emerging as judges for precise multimodal evaluations, yet existing benchmarks organized by task types overlook core judgment capabilities...

• VideoChat3 delivers a fully open 4B video MLLM that outperforms prior open-source models on general, long-form, and streaming benchmarks.
  -...

Moonshot AI's 2.8T-parameter Kimi K3 shows open-weight models have reached parity with US frontier systems, topping benchmarks in coding and browsing...

VideoChat3 introduces a fully open video-centric MLLM with 4B parameters that delivers strong generalization across general, long-form, and streaming...

Existing MCP-based benchmarks overlook continuous tool interface evolution, producing flawed evaluations of LLM agent adaptability. MCPEvol-Bench...

GPT-5.6 Luna leads the overall aggregate (81 vs 76) across 20 shared benchmarks, with big edges in knowledge (92.3 vs 60.4) and agentic tasks like...

Sol's major gains in object detection and counting turn OpenAI's highlighted UI agents and detailed 3D visualizations from demos into practical tools....

Thinking Machines Lab's Inkling arrives as a 975B MoE open-weights model optimized for enterprise customization rather than raw leaderboard...

Moonshot's Kimi K3 arrives as the largest open-weights model at 2.8 trillion parameters, set for public release July 27. Its benchmarks place it just...

Boogu-Image-0.1 is a competitive Apache-2.0 open-source model family (Base, Turbo, Edit) for unified image generation and editing that matches closed-source systems using only 208M images and ~$400K compute.

GPT-Red's self-play RL training generated diverse prompt injection attacks that trained GPT-5.6 Sol, slashing its vulnerability rates to ~3.8% for...

• Inkling (975B total/41B active) is Thinking Machines' first open-weights multimodal MoE, pretrained on 45T tokens across text/images/audio/video...

Google has confirmed Gemini 3.5 Pro is internally deployed as the successor to Gemini 3.1 Pro, positioned above the May 2026 Gemini 3.5 Flash release...

Thinking Machines Lab's Inkling release supplies Western enterprises a US-developed alternative to dominant Chinese open-weight models like DeepSeek...

Two major world models for physical AI emerged this week, signaling accelerating convergence of foundation models and robotics.

• Xiaomi...

The AI community is highlighting Thinking Machines' restrained, detail-rich launch of Inkling as a refreshing contrast to typical hype cycles.

-...