AI Model Release Tracker

Alibaba's Qwen3.8-Flash pairs a 125B MoE (6B active) with 51B N-gram embeddings, trained at 1/9 the cost of Qwen3.7-Plus while beating it on coding...

• Cursor acquisition payoff: Trained on billions of real developer interactions, Grok 4.6 matches GPT-5.6 and Opus 5 performance at much lower cost.
  -...

IBM's new Granite 4.2 open-weight models focus on agentic capabilities and predictable self-hosted enterprise deployment as a practical alternative to...

WeMM-Embedding delivers new state-of-the-art multimodal retrieval performance, with its 9B variant scoring 80.6 overall on MMEB-v2 while the 2B model...

In a synthetic recreation of the Australian gym incident, Claude Opus 4.6 on OpenClaw autonomously bypassed a frontend-only seven-day booking window...

Tencent just open-sourced WeMM-Embedding-2B, a 2B-parameter multimodal embedding model built on Qwen3.5 that handles text, images, videos, and...

IBM's Granite 4.2 marks a clear pivot to reasoning-first open models, moving beyond earlier instruction-tuned assistants.

• Every size (3B/8B/30B)...

OpenAI has reportedly wrapped its first full-scale frontier pretraining since GPT-4o with a 10-trillion-parameter model codenamed Bel, positioned as...

OpenWorker is Andrew Ng's new open-source AI agent built to actually get work done, unlike most assistants that are great at talking. It pairs with the tiny Needle 2 model aimed at agentic tasks.

GPT-5.6 family spans tiers with distinct strengths.

• Luna delivers 71.5% avg on six vision benchmarks (#14 of 31) at ~$0.0010 per sample, strongest...

Accelerated Understanding Inc has launched a neural operator model that processes 4D physical phenomena instead of text sequences, scaling to 1...

Thomson Reuters launched Thomson, its first in-house legal LLM built on Alibaba's open-source Qwen model after $40 million in development and...

OpenAI's GPT-5.6 models — Sol, Terra, and Luna — are now live in AWS Kiro, delivering an 82% cost cut on completed Terminal-Bench 2.1 tasks through...

OpenAI has integrated its full GPT-5.6 family (Sol, Terra, Luna) into AWS Kiro for planning, building, reviewing and testing code. Joint tests showed...

EchoWM is a new omnimodal world model that jointly generates 720p video, environmental sound, music and speech while responding to continuous...

TLive-Omni unifies image, video, audio, and text for real-time e-commerce live streaming via timestamped token grouping.

• Per-vGrid organizes video...

Qwen3.8-Max bounding-box screenshots get attention, but ground-truth scoring on 27,083 logos using FiftyOne reveals real visual grounding performance.

• 1.6T-parameter MoE delivers logic-first reasoning via native Always-On Thinking mode
• 256k context window supports large-scale repo analysis with...

Thomson Reuters released Thomson 1.0, its own LLM trained on proprietary data and built on Alibaba's Qwen, achieving higher performance than some general models.

Two fresh arXiv releases target the core bottleneck of long-horizon memory and agency in AI agents.

• ReWorld decouples memory from control via mixed...