Google releases Gemini 3.1 Flash Live, claims improved audio recognition and lower latency for voice conversations

Google has released Gemini 3.1 Flash Live, its highest-quality audio model designed for natural and reliable real-time voice interactions. The model scores 90.8% on ComplexFuncBench Audio and 36.1% on Scale AI's Audio MultiChallenge with thinking enabled. It's now available to developers via the Gemini Live API, enterprises through Gemini Enterprise for Customer Experience, and consumers in Search Live and Gemini Live across 200+ countries.

ByteDance confirmed Thursday that Dreamina Seedance 2.0, its audio and video generation model, is rolling out in CapCut across seven initial markets. The model generates videos up to 15 seconds with realistic textures and motion, but includes safety restrictions blocking generation from real faces and unauthorized IP use.

Google has released Gemini 3.1 Flash Live, its highest-quality audio and voice model designed for real-time dialogue. The model scores 90.8% on ComplexFuncBench Audio and 36.1% on Scale AI's Audio MultiChallenge with reasoning enabled, with improved tonal understanding and lower latency compared to previous versions.

GitHub will use Copilot interaction data from Free, Pro, and Pro+ plan users to train AI models starting April 24, 2026, unless users actively opt out. The policy does not affect Copilot Business and Enterprise customers. Data shared will include prompts, outputs, code snippets, filenames, and repository structures.

Google unveiled TurboQuant, a compression technique that reduces memory required to run large language models by six times by optimizing key-value cache storage. Memory chipmakers Samsung, SK Hynix, and Micron fell 5-6% on concern the efficiency breakthrough could reduce future chip demand. Analysts expect the decline reflects profit-taking rather than a fundamental shift, as more powerful models will eventually require more advanced hardware.

The ARC Prize Foundation released ARC-AGI-3, an interactive benchmark requiring AI agents to explore environments, form hypotheses, and execute plans without instructions. All 135 environments were solved by untrained humans, yet frontier models—including Gemini 3.1 Pro Preview (0.37%), GPT 5.4 (0.26%), Opus 4.6 (0.25%), and Grok-4.20 (0.00%)—scored below 1%.

Mistral AI released Voxtral TTS, an open-source text-to-speech model designed for enterprise voice agents and edge devices. The model supports nine languages, adapts custom voices from samples under five seconds, and achieves 90ms time-to-first-audio latency with a 6x real-time factor.

Apple researchers developed RubiCap, a framework for training dense image captioning models that achieve state-of-the-art results at 2B, 3B, and 7B parameter scales. The 7B model outperforms models up to 72 billion parameters on multiple benchmarks including CapArena and CaptionQA, while the 3B variant matches larger 32B models, suggesting efficient dense captioning doesn't require massive scale.

Google Research unveiled TurboQuant, a lossless memory compression algorithm that reduces AI inference working memory (KV cache) by at least 6x without impacting model performance. The technology uses vector quantization methods called PolarQuant and an optimization technique called QJL. Findings will be presented at ICLR 2026.

Google has released Lyria 3 Pro, its latest AI music generation model capable of creating tracks up to three minutes long with improved understanding of musical structure. The model is available through Gemini, Google Vids, Vertex AI, and Google AI Studio. Google claims the training data comes from sources it has contractual and legal rights to use.

Amazon Bedrock has introduced three distinct video analysis workflows that leverage multimodal foundation models to extract insights from video content at scale. The approaches—frame-based, shot-based, and multimodal embedding—are designed for different use cases and cost-performance trade-offs, with open-source reference implementations available on GitHub.

Amazon Bedrock now enables reinforcement fine-tuning (RFT) across multiple model families including Amazon Nova, open-weight models like OpenAI's GPT-OSS 20B, and Qwen 3 32B. The service automates the end-to-end customization workflow using GRPO optimization, allowing models to learn from feedback on multiple responses rather than static training datasets, with support for OpenAI-compatible APIs.