Model Releases

Meta has released Muse Wonder and Muse Wonder Full, marking its first frontier-model launch from the Superintelligence Lab. The models target harder tasks such as coding while putting long-term user memory at the center of the Muse assistant experience.

PolyOCR-Venus brings recognition, localization, document parsing, information extraction, and OCR reasoning into a shared instruction-following framework. Its training method selectively uses teacher supervision, while OCRBench v2.1 updates the evaluation with manually checked annotations and task-aligned scoring.

Reflection AI has introduced Beam, its first frontier open-weight model, claiming performance comparable to leading Chinese open models with substantially lower inference requirements. The startup is targeting enterprises and sovereign institutions with a broader “AI factory” strategy.

Google Research introduces TabFM, a 400-million-parameter foundation model for tabular prediction. It treats supervised table learning as in-context learning and produces predictions without task-specific fine-tuning.

Hunyuan-A13B is an open-source large language model built with a Mixture-of-Experts architecture. It has 80 billion total parameters but activates only about 13 billion during inference, while using adaptive reasoning to balance quality and latency.

Anthropic’s Opus 5.5 and OpenAI’s GPT-6 Sol and Luna focus less on dramatic new capabilities than on lower prices, faster output, and practical deployment. The competition is increasingly about value, not just raw performance.

xAI has announced Grok 4.7 as a model for coding and knowledge work, with an emphasis on multi-hour tasks, self-checking, and safeguards. Its published performance figures are company-reported and still require independent validation.

TypeSafe AI has introduced Jev, its first “System One Model,” and opened it to early access. Rather than focusing on conversation, Jev is designed around turning unstructured state into fast, structured decisions that software can use.

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has launched Jev, a model that returns calibrated probabilities instead of text. Early developers are using it for classification, routing, and software safety at lower cost and latency.

LimiX-2 introduces Contextual Mechanism Networks, a framework that models the joint structure of variables rather than focusing only on a designated target column. The model reports leading results on TabArena, TALENT, and BCCO while extending tabular foundation models toward causal skeleton recovery.

Suno’s v6 family is the company’s first music model developed with support from the record industry. It improves genre matching, editing, and multimodal prompting, but still struggles to reproduce the deliberate irregularities of human performance.

OpenAI began rolling out GPT-6 Astra but left many paying users unable to use the model on launch day. CEO Sam Altman apologized for the “messy rollout,” while the company still has not offered a firm timeline for wider access.

OpenAI has launched GPT-6 Astra, highlighting major gains in cybersecurity, software engineering, science, and agentic computer use. The release also raises difficult questions about oversight after recent reports of model-driven attacks.

WeatherNext 3 combines live satellite observations, hourly refreshes and finer spatial grids to improve forecasts for fast-changing and localized weather. It also adds variables designed for wind and solar energy planning.

Google has launched Gemini 3.8 Flash, a model designed to take more reasoning steps and call tools repeatedly on difficult tasks. Its per-token price is unchanged, but heavier usage could make individual tasks more expensive.

Google DeepMind has introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, emphasizing longer software-engineering tasks, complex reasoning, vulnerability discovery, and automated patching. The models retain Flash’s speed and pricing position while using more reasoning and tool calls on difficult tasks.

Anthropic says Claude Fable 5.1 improves on its predecessor while costing about 25 percent less in typical use and up to 45 percent less on complex agentic workloads. The release also revises safeguards and outlines a new enterprise data-retention option.

Tencent Hunyuan has introduced Hy4 preview, an open model with 770B total parameters, 4.9B activated parameters, and a 1M-token context window. The release also highlights integrations with game engines and MCP-oriented tool workflows.

Alibaba’s DAMO Academy and clinical partners have developed DAMO LiON, an AI system that assists doctors in detecting small malignant liver lesions on contrast-enhanced CT scans. In a two-month prospective real-world trial, it helped identify 15 previously overlooked malignancies.

Ox Alpha has drawn intense attention since its free release on OpenRouter, yet the team behind it remains anonymous. Speculation has ranged from Z.ai’s GLM models to an unreleased Microsoft MAI system, with no confirmed answer.

Motif 3 is a decoder-only sparse MoE language model with 314B total parameters and 13.2B activated per token. Its design combines GDLA, training-stability tricks, and multi-teacher post-training to improve reasoning, coding, and long-context performance.