Mistral opened a Large 4 API preview ahead of its promised weight release

TOP 3

1. Mistral made Large 4 available through a preview API while promising weights later.
2. Google released EmbeddingGemma 2 for local cross-modal retrieval.
3. Anthropic expanded its Cyber Verification Program into tiered access.

Models

• Mistral Large 4 preview: Mistral’s October 6 announcement says its natively multimodal mixture-of-experts model has 1 trillion total parameters and 49 billion active parameters. Developers can use the public preview API in Mistral Studio now, but the weights are not yet available; Mistral says it will release them by the end of the month after red-teaming with selected partners. The company reports strong coding, cyber, finance and visual-grounding results, including on third-party benchmarks, but the comparative claims in this announcement are its own account rather than independent verification here.
  • What remains open: Mistral says more architecture details, additional benchmarks and post-training methods will accompany the weights. This is therefore an API preview, not an open-weight download or a complete technical report. It follows yesterday’s Reflection Beam preview, but leads today because the API is usable now and Mistral documents multimodal and enterprise workloads rather than only announcing future artifacts.
• EmbeddingGemma 2: Google’s October 6 release describes a 740-million-parameter model that embeds text, code, images, audio and video into one space under Apache 2.0. Google says the weights are downloadable on Hugging Face and Kaggle, with optional vision and audio encoders; its quality and device-memory comparisons are Google-reported.
• Falcon OCR Arabic: Technology Innovation Institute’s October 6 account describes adapting its 270M-parameter OCR model to Arabic documents with supervised fine-tuning and reinforcement learning. On the institute’s own Arabic-document evaluation, it reports 81.9% text accuracy, behind Gemini 3.5 Flash’s 84.3%; the page’s linked model reference points to the base Falcon-OCR repository, so this post does not establish a separately downloadable Arabic checkpoint.

Tools & Open Source

• Cyber access tiers: Anthropic’s October 6 announcement folds Project Glasswing into an expanded Cyber Verification Program with Defense, Red Team and Specialized access, each with different verification and safeguards for security work. Anthropic says qualifying applicants can seek access to Opus, Sonnet and Mythos models with reduced cyber blocking; the program requires verification and is not a general removal of restrictions.

Community

• Discussion, not a model evaluation: In an October 6 Hacker News thread about EmbeddingGemma 2, Simon Willison argues that openly licensed embedding weights provide a fallback if a hosted provider discontinues a model and users would otherwise have to recompute stored vectors. That is a developer’s portability concern, not evidence for Google’s performance claims.