EmbeddingGemma 2 Brings Multimodal Search to Devices
Google DeepMind launched EmbeddingGemma 2, an open 740M model for on-device vector search across text, code, audio, and video under Apache 2.0.
Models
Model cards, releases, evaluations, and the limits a lab states in public.
Anthropic Expands Cyber Verification Program Into 3 Tiers
Anthropic structured its Cyber Verification Program into three tiers, easing safety filters on Claude models for vetted defenders and security auditors.
DeepMind Validates SynthID Bio Watermark in Living Systems
Google DeepMind embeds SynthID Bio watermarks into generative proteins and DNA, keeping biological function intact while leaving fraud defenses open.
Solar Mini 4 Brings 3B MoE and 512K Context to Nous Portal
Upstage's Solar Mini 4 pairs 3B active parameters with a 512K context window, launching with two weeks of free access on the Nous Portal platform.
OpenAI and Cerebras: Altman Reaffirms Chip Deal
Sam Altman dismissed market rumors and confirmed OpenAI and Cerebras continue working on fast inference, boosting the chipmaker's stock.
AI token gray market dodges US chip sanctions
Chinese brokers use shell firms and stolen credentials to resell Claude tokens at 90% discounts, fueling domestic AI model distillation.
Reflection AI Beam: New 501B Open-Weight Model
Reflection AI announced Beam, a 501B open-weight model trained on 10,500 GPUs to challenge Chinese labs. Public weights arrive by late October 2026.
Perplexity open-weight models target routing and search
Perplexity released open-weight models pplx-decider and pplx-embed to cut costs for agent workflows and contextual vector search systems.
Fine-Tuning Nemotron 3.5 ASR for Regional Accents
NVIDIA published an open guide to adapt Nemotron 3.5 ASR to regional dialects in 4.5 hours while avoiding catastrophic forgetting.
NVIDIA Speeds Up GPT-6 Astra Ultrafast on Blackwell
NVIDIA and OpenAI boost GPT-6 Astra Ultrafast on Blackwell GPUs. Output speeds jump 8x, but overall latency drops 2.6x at $300 per million tokens.