AI Model Launch Tracker

What we captured

NVIDIA released the open-weight Nemotron 3.5 Super VL model for visual and language tasks, and made its weights available under the OpenMDW license.

• • Nemotron 3.5 Super VL has 120 billion parameters.
• • The model is designed to work across PDFs, charts, tables, images, and webpages, and to use tools.
• • NVIDIA says the OpenMDW license permits commercial use, fine-tuning, and redistribution.
• • NVIDIA reports partner post-training gains of 9% to 44% across five specialized evaluations, but does not provide task definitions, sample sizes, or testing protocols.

Source evidence

Confidence 98% / Completeness 91%

What we captured

Black Forest Labs released the 5.6-billion-parameter open-weights image model FLUX 3 Image Klein.

• • Released as a 5.6-billion-parameter open-weights model for image generation and editing.
• • Supports native 4K output and up to ten reference images.
• • An INT8 build runs on NVIDIA GPUs with 8 GB of VRAM.
• • BFL API pricing is $0.009 per 1,000 images.
• • The stated license permits free commercial use for companies with annual revenue below $5 million.

Source evidence

What we captured

Nvidia released open weights for Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model for enterprise agents and document analysis.

• • The model has 120 billion parameters, with 12 billion active per token, according to Nvidia.
• • Nvidia released base, post-trained, and 4-bit compressed weights.
• • Datasets, training recipes, and reinforcement-learning environments were also published.

Source evidence

Confidence 99% / Completeness 87%

What we captured

Liquid AI released an open-weight reasoning model for local agent workflows, with 24B total parameters, 2.45B active parameters per token, and a 131,072-token context.

• • The model activates 2.45B parameters per token.
• • Liquid AI reports an AutomationBench score of 27.69, versus 8.72 for Qwen3.6-35B-A3B.
• • Liquid AI reports a SWE-Bench Verified score of 51.00, versus 66.20 for Qwen3.6-35B-A3B.
• • Listed API prices are $0.10 per million input tokens and $1.00 per million output tokens.

Source evidence

What we captured

Fish Audio released Drama 3's full weights on Hugging Face under the Fish Audio Community License v1.0.

• • Fish Audio published Drama 3's full weights on Hugging Face.
• • The license permits free commercial use for companies and affiliates earning less than $500,000 annually.
• • Artificial Analysis reports a 93.2% score on Pronunciation Robustness.

Source evidence

Confidence 97% / Completeness 91%

What we captured

Neura Robotics announced Neuron AI, a multimodal foundation model for physical AI combining vision, touch, force, audio, robot state, language, and memory.

• • Announced October 14, 2026.
• • Combines 3D vision, touch, force, audio, robot state, and language with memory.
• • Training uses real-world robot data, simulated data, and human demonstrations.
• • Neura says training spans five of its robot embodiments.

Source evidence

What we captured

Tether released the 254-million-parameter QVAC TranscribePsy model with open weights for local speech transcription.

• • The 254-million-parameter model is designed for local speech transcription without cloud access.
• • Tether offers versions supporting 25 European languages and a smaller eight-language configuration.
• • Tether reported 1.65-second transcription of 10 minutes of audio on an iPhone 17 Pro Max using Core ML.
• • The model weights and evaluation configurations are available under Apache 2.0.

Source evidence

Confidence 99% / Completeness 91%

What we captured

xAI announced Grok Voice 3.0, a speech-to-speech model available in the Voice Agent API at $0.08 per minute.

• • Announced October 14, 2026.
• • Available in the Voice Agent API at $0.08 per minute, the same price as Grok Voice 2.0.
• • Artificial Analysis reported a Speech-to-Speech Index score of 86.9.

Source evidence

Confidence 99% / Completeness 87%

What we captured

Paper Instruments announced Ultramarine 1 Flash for professional work; the company said model weights were expected in about two weeks.

• • CEO Daanish Khazi announced the model on October 14, 2026.
• • Khazi said a five-person team post-trained it for about $300,000.
• • The company reported a $0.13 cost for one task in its internal benchmark.
• • Model weights were expected in about two weeks.

Source evidence

Confidence 94% / Completeness 87%

What we captured

Tether released QVAC TranscribePsy, a 254-million-parameter multilingual speech-recognition model with on-device builds and open weights under Apache 2.0.

• • 254 million parameters.
• • Laptop version supports 25 European languages; mobile version supports eight.
• • Includes open weights and mobile builds under Apache 2.0.
• • Tether reports 1.65-second transcription of 10 minutes of audio on an iPhone 17 Pro Max.

Source evidence

Confidence 99% / Completeness 91%