Third-party technical analysis on June 18 confirmed the benchmark claims for MiniMax M3, the open-weight model released June 1, finding the MiniMax Sparse Attention (MSA) architecture’s reported performance figures credible (TechTimes). The model carries roughly 428 billion total parameters with approximately 23 billion activated per forward pass (datanorth.ai), and the MSA mechanism delivers approximately 15.6x faster decoding and 9.7x faster prefill at 1 million tokens compared to its predecessor M2 (felloai). On the SWE-Bench Pro software-engineering benchmark, M3 scores 59.0%, placing it above GPT-5.5 and Gemini 3.1 Pro on that task (nerova.ai). The weights reached Hugging Face on June 7 and the MSA technical report was posted to arXiv on June 11 (The Decoder). API pricing stands at $0.60 per million input tokens and $2.40 per million output tokens; commercial deployments require a separate license agreement under the MiniMax Community License (felloai).
xAI’s Grok 4.3 became generally available on Amazon Bedrock on June 16, adding a 1 million-token context window and configurable reasoning effort levels - none, low, medium, or high - to AWS’s managed model catalog (AWS). The model runs on Mantle, a new Bedrock inference engine designed for price-to-performance efficiency, with support for tool calling, structured outputs, and streaming responses (Artificial Analysis). On June 18, Anthropic opened a Seoul office and signed a memorandum of understanding with South Korea’s Ministry of Science and ICT to collaborate on AI safety evaluation in Korean-language contexts and exchange information on AI-enabled cyberthreats (Anthropic); enterprise partners named at launch include Samsung SDS, LG CNS, and Hanwha Solutions, all deploying Claude at scale across their organizations (UPI).