Prime Intellect’s Prime Mega MoE makes DeepSeek’s fused MoE kernel trainable, cutting Qwen3 training steps up to 55% ...

Block’s AI research lab has released Dessa, a full-duplex speech model trained only on human conversation. Block announced it on October 23, 2026 as a research preview. Dessa is a speech-to-speech ...

AssemblyAI-TTS-1 is a streaming text-to-speech API for voice agents, priced at $15 per 1M characters with unlimited ...

It is the second 0.6B release in Edge0’s Audio8 TTS family. It builds on Audio8-TTS-Preview-0.6b and keeps the same ...

Edge0 released Audio8-TTS-Realistic, a 0.6B open-weight zero-shot voice cloning TTS model under Apache 2.0. It supports 23 languages and adds a final GRPO stage driven by a learned realism critic.

Baseten has released joey-1, a 2B open-weight decision model that returns calibrated probabilities instead of text. It runs at 20 ms P50 latency and costs $0.01 per million input tokens.

Infini-AI-Lab released Hermit, an Apache-2.0 harness adapter framework that trains an external layer of instructions, skills, memory and tools from task feedback, leaving the agent’s model and harness ...

Gemini 4 Carbon, Google DeepMind’s frontier model, ranks #1 on Artificial Analysis Index at 59 and leads cybersecurity.

MiniMax open-sources MCode Stack: 15 agent skills and 5 checklists its MiniMax Code team uses for harness engineering ...

FlashMap speculative decoding uses a 0.5B flow map drafter to reach 6.14x lossless speedup on Qwen3-8B, beating DFlash.

Google DeepMind’s Genie 4 world model generates interactive worlds with native audio, dialogue, and Gemini 3.8 Flash planning.