
Zyphra's Zamba2-1.2B Beats Gemma2-2B While Using 6x Less Memory
Zyphra's 1.2B hybrid SSM-Transformer model beats Gemma2-2B on MT-Bench while using 6x less KV cache memory
20 stories

Zyphra's 1.2B hybrid SSM-Transformer model beats Gemma2-2B on MT-Bench while using 6x less KV cache memory

Zyphra's Zamba2-7B-Instruct blends Mamba2 and shared attention to beat Llama3-8B on speed and memory while matching it on quality

Zyphra releases Zonos-v0.1, a 1.6B open-weight TTS model with voice cloning, emotion control, and 44kHz output that rivals ElevenLabs — free under Apache 2.0

Ant Group's Robbyant open-sourced a 1B-parameter ViT backbone that treats object boundaries as a pretraining signal, matching 7B DINOv3 on depth benchmarks.

Aiden is a fully local, open-source AI agent for Windows and Linux that controls your computer, runs 1500+ skills, and never sends data to the cloud
Google's ScientistOne automates the full research pipeline — literature review, experimentation, and paper writing — while being the only system to achieve zero hallucinated references across a 75-paper audit.
A 7.7M active-parameter multilingual embedding model that matches or beats models 3-40x its size on retrieval, running on a Raspberry Pi.
Google DeepMind's SURF learns to separate mixed audio without ever seeing clean isolated sources, setting a new unsupervised state-of-the-art at ICML 2026
Google's Project Astra research prototype makes its ICML 2026 debut, showcasing a universal AI assistant that sees, hears, reasons, and acts in real time

Artificial Analysis launches six domain-specific AI benchmarks grounded in real job tasks, revealing sharp tradeoffs between frontier performance and cost.

Gemini Spark gains real-time event tracking, macOS support, MCP integration, and new third-party app connections in a major update to Google's 24/7 AI agent

NVIDIA quantizes Moonshot AI's 1T-parameter Kimi-K2.7-Code to FP4, making it faster and cheaper to run on Blackwell GPUs without sacrificing accuracy.

GPT-Realtime-2.1-mini brings reasoning and tool use to the budget voice tier, plus a 25%+ p95 latency cut across all Realtime voice models

Factory and Hugging Face team up to let developers upload Droid agent traces directly to the Hub, fueling an open dataset push for next-gen coding models.

Genkit's new Agents API bundles message history, streaming, persistence, and human-in-the-loop into one interface for TypeScript and Go

prime-rl's new algorithms layer ships six built-in RL methods and lets you mix training signals per environment in a single run

Google connects Gemini directly to Google Business Profile, giving small business owners an AI assistant that remembers their brand, analyzes their performance data, and can update their listing in real time.

Anthropic discovers a hidden mental workspace inside Claude that reveals silent reasoning, exposes deceptive intent, and can be shaped to improve model honesty

SGLang integrates DSpark's confidence-driven speculative decoding, delivering best-in-class throughput/latency tradeoffs for DeepSeek-V4 at any batch size

Researchers from Meta, Google DeepMind, Cornell, and NVIDIA put a hard number on LLM memorization: ~3.6 bits per parameter, with a new framework that cleanly separates memorization from generalization.