Which AI writes the better take? You decide — blind.

A 45M-parameter agentic LLM has been compressed to a 14MB binary, enabling it to run on devices with as little as 28MB of RAM, such as sub-$200 phones and microcontrollers, at speeds of up to 500 tokens/sec, allowing for on-device AI capabilities like tool calling and structured extraction without relying on cloud processing.

NVIDIA released Magpie TTS Multilingual, a 364M-parameter open-weights text-to-speech model covering 12 languages (now including Arabic, Korean, and Brazilian Portuguese) with male/female voices, deployable via NIM on your own infrastructure. Because it's the final and most latency-sensitive stage of a voice pipeline, running it yourself lets you tune latency, enforce data residency, and fine-tune per domain rather than being locked into a black-box audio-in/audio-out API — worth evaluating if you're building cascaded ASR+LLM+TTS agents and want control over the last mile users actually hear.

OpenAI split Daybreak into two tiers, with the new GPT-5.6-Cyber model (built on GPT-5.6 Sol) gated behind the Red tier and initially limited to trusted partners like Crowdstrike, Cloudflare, IBM, and Accenture for offensive-style vulnerability research and security testing. The practical takeaway: frontier cyber capability is now productized and access-controlled, so if you're not one of the anointed partners you get the defensive Blue tier (incident response, malware analysis, patch validation)—and you should assume adversaries are already running comparable autonomous tooling against your stack, narrowing your prep window.

Meta shipped Muse Glimmer, a dense 30B Apache-2.0 multimodal VLM distilled from their larger Muse model, with day-0 support in transformers, llama.cpp, vLLM, and Inference Endpoints, plus an optional speculative-decoding drafter that speeds up structured/code generation at a memory cost. This is a genuinely local-deployable agentic model with a permissive license and multimodal tool calling, so you can run privacy-sensitive coding assistants, document analysis, and Claw/Hermes-style agents on-prem without API costs or vendor lock-in—and the 2B Perception Encoder means image/video handling is a first-class capability, not a bolt-on.

Meta has released a 30-billion parameter AI model, Glimmer, that can run AI agents locally on consumer hardware with a single GPU, enabling tasks like managing schedules and drafting messages without cloud processing, and this shift to local processing on consumer devices improves privacy for personal agent applications.