This Week's New AI Models: Robots, Voice, Genomics and Image Editing

What this is about

The Korean site AI Times rounded up eight new AI models in a roundup dated October 23, 2026. They range from robot control to speech AI, genomics and image editing. All performance figures come from the vendors.

Dream-1: controlling robots the model has never seen

Dream Labs is coming out of stealth with Dream-1. It's a "robot foundation model," or "world action model," designed to control dual-arm robots with parallel grippers without extra training or demonstration data. Typical tasks include moving objects or closing drawers. Access is via API from $3 per hour. According to the company, performance depends on the robot's setup and control software.

Ming-3.0: open audio models from Ant Group

InclusionAI, Ant Group's research group, is releasing five models in three families:

Audio Flash understands long recordings, music and ambient sounds, mainly in Chinese and English, and replies by voice. The recognition model can use technical terms and earlier conversation history. "Active parameters" means that in Mixture-of-Experts models, only a portion of the total weights is used at each step.

Eleven Realtime and other speech models

ElevenLabs is releasing two speech-to-speech models for real-time conversations. They coordinate speech recognition, turn-taking, the language model and speech synthesis, so developers don't have to wire these pieces together themselves. v1 Mini is for low-cost everyday dialogue, v1 Max for complex instructions and multi-step tool use. Both support over 90 languages and over 17,000 voices, and offer an interface compatible with the OpenAI Realtime API. ElevenLabs says it responds within half a second.

Bark 3.0 (Underdog) is an English speech recognition model with open weights. It's 163 MB and can run locally. Averaged across seven English test sets, its word error rate is 4.94%. On an Apple M2, one minute of audio takes about 1.5 seconds. Gradium's Phonon is a family of small text-to-speech models under 50M parameters, built for phone and laptop CPUs and for browsers. Phonon-31m-en has 31M parameters (excluding the audio tokenizer) and is 64.5 MB at 8-bit. It produces 24 kHz speech and can stream. It's released under Apache 2.0; the 40M models are non-commercial only.

Design, genomics and image editing

CocoLabs' Vyb creates marketing graphics using a brand's logos, fonts and product photos. It outputs layers rather than a flat image, so titles or photos stay editable afterward. It's in closed beta with a waitlist.

Radical Numerics' Omnii-1 is a multimodal genome language model trained on 28 data types, such as DNA, RNA, proteins and gene expression. It was further trained as a research preview for eight applications, including genome interpretation, disease detection, antibody design and proteins for rare earths. According to a post cited in the roundup, it reportedly designed AI-generated bacteriophages that were used on a patient with an antibiotic-resistant infection. The source gives no further evidence for this.

Bria's Fibo 2.0 combines image generation and editing, was trained only on licensed data, and improves text rendering, textures and depictions of people. On Artificial Analysis, it ranks first among open models for image editing. The weights are available for non-commercial use only.

Our take

The trend: speech AI is getting smaller and more local (Bark, Phonon), while also being offered as a ready-made real-time agent (Eleven). Dream-1 and Omnii-1 in particular make ambitious claims that need independent verification. Benchmarks are vendor-reported, and for several models, licenses restrict commercial use.