Moonshot AI

Telnyx Voice Agent Switches Languages Mid-Call

A Telnyx developer example demonstrates a voice AI assistant that detects and follows a caller's spoken language on every turn, enabling real-time code-switching within a single call. The example is a minimal Python Flask app that uses Deepgram's nova-3 STT with language:auto for automatic language detection, an LLM instructed to reply in the detected language, and a single multilingual TTS voice (voice ultra katie). The demo supports five languages (English, Spanish, Portuguese, Hindi, Mandarin) and is provided as an open-source code example in Telnyx's GitHub repository. The conversation and language handling run entirely on Telnyx's assistant without application-layer language routing or separate interpreter services.

Microsoft directs staff to default to OpenAI GPT-5.6 Sol

Microsoft has instructed engineers working on AI coding projects to default to OpenAI’s flagship model, GPT-5.6 Sol, inside GitHub Copilot as part of an internal efficiency push. The guidance came in a memo from Jay Parikh, executive vice president of Microsoft’s CoreAI engineering group, who said shifting more workloads to OpenAI models helps Microsoft get greater value from its token spending. The company continues to offer other model options and said it periodically updates internal default model settings. The move occurs alongside Microsoft’s ongoing commercial relationships and investments across the AI ecosystem, including ties to Anthropic and prior structural agreements with OpenAI.

Technical Primer on Kimi K3 and KDA Inference Kernels

This SemiAnalysis technical primer explains the core techniques behind the Kimi K3 model architecture, focusing on Kimi Delta Attention (KDA), its ancestry (linear attention, DeltaNet, Gated DeltaNet), and system-level implementations such as FlashKDA. The article describes FlashKDA’s kernel design (K1/K2), arithmetic and memory complexity (O(T*D^2) compute at chunk level, prefill linear, decode constant), interactions between KDA and full-attention MLA in hybrid architectures, attention-residual and block-attention-residual designs, LatentMoE communication trade-offs, and load-balancing via Quantile Balancing. The piece also reports benchmarking activity (InferenceX), day‑0 bringup notes (NVIDIA/AMD recipes on vLLM), and observed OpenRouter provider pricing floors as of July 30.