Mistral Large 4 enters public preview as a 1-trillion-parameter open-weight model
Mistral released Large 4, a natively multimodal hybrid instruct-and-reasoning mixture-of-experts model with 1T total and 49B active parameters. It is in preview on Mistral Studio at $1.36 per million input tokens and $4.18 per million output tokens; Mistral says the weights follow by the end of October, so it is announced and available via API but not yet downloadable. Mistral's own benchmark claims centre on cybersecurity (a top-five placement on the Artificial Analysis Cyber Index, 93% on Cybench), coding (61.7% on DeepSWE v1.1), and finance/legal agent tests where it says it beats GPT-6 Astra. These figures are vendor-reported and not yet independently reproduced; the context window was not stated on the announcement page.
Sources & depth
Google releases EmbeddingGemma 2, an Apache 2.0 embedder for text, images, audio and video
EmbeddingGemma 2 maps text, code, images, video and audio into one shared embedding space, so a single index can serve cross-modal search. It is 740M parameters in total but modular: text-only needs 270M, with optional vision (+170M) and audio (+300M) encoders. Vectors are 768-dimensional and can be truncated to 512, 256 or 128. Google reports MTEB Code rising from 68.76 to 78.68 and leading the sub-1B multimodal embedders. Weights are on Hugging Face and Kaggle, and support landed the same day in Transformers v5.19.0, Unsloth, Ollama and others.
Sources & depth
Ollama v0.40.0 makes MLX the default runtime on Apple Silicon
Ollama's v0.40.0 release runs supported models on Apple's MLX framework by default on Apple Silicon Macs. MLX support covers Qwen3.8, Gemma 4, Qwen3.6 and Qwen3.5, and the release adds EmbeddingGemma 2 embeddings (including per-item image and audio input through the embed API) on the MLX runner. It follows a run of release candidates over the past week.
Sources & depth
Anthropic merges its cyber access programs into one three-tier Cyber Verification Program
Anthropic combined Project Glasswing and the original Cyber Verification Program into a single offering with three tiers: Defense Access (incident response, malware analysis, vulnerability validation; approval in a few days), Red Team Access (authorized penetration testing; a few weeks, organizations only for now), and Specialized Access (verified organizations testing critical safety systems such as power grids, reviewed together with the US government). All tiers include Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, the last of which is otherwise limited-availability. Defense Access is open to companies, nonprofits, universities, open-source maintainers and individual researchers with a disclosure history.
Sources & depth