AI Model Launch Tracker

What we captured

Meta launched Muse Voice Transcribe, a real-time multilingual speech-to-text model with speaker diarization, code-switching, and long-audio support.

• • Announced September 1, 2026.
• • Supports real-time streaming transcription, more than 20-speaker diarization, code-switching, and audio inputs exceeding one hour.
• • Trained on more than 70 languages, with 25 validated at launch.

Source evidence

Confidence 98% / Completeness 87%

What we captured

Runway announced Solaris, a research-stage Interface World Model that generates and dynamically updates software interfaces in response to user interaction.

• • Solaris generates interfaces frame by frame rather than executing conventional interface code.
• • The system responds to clicks, drags, and voice commands and renders at 720p.
• • Runway says text rendering, reliability, long sessions, and screen-reader compatibility remain unresolved.

Source evidence

What we captured

Google Research introduced MAPL-EMIT, a Swin-S-based deep-learning model for detecting, quantifying, and localizing methane plumes in NASA EMIT satellite data, and released its trained model and related datasets and tools.

• • MAPL-EMIT was introduced in a Google Research post dated September 1, 2026.
• • The model was developed in collaboration with NASA JPL for global methane point-source monitoring using EMIT data.
• • The release includes a global plume database, trained model, synthetic plumes, and an inference library.

Source evidence

Confidence 97% / Completeness 87%

What we captured

Visko opened public access to Orbis 1.0, a real-time interactive video foundation model that streams continuously generated video.

• • Streams interactive video continuously in real time rather than rendering a fixed clip.
• • Visko reports 4K output at 24 frames per second and hour-scale generation without evident visual or color drift.
• • Users can change prompts during generation, with updates reflected while streaming continues.

Source evidence

Confidence 98% / Completeness 87%

What we captured

Skild AI announced S1, a robot foundation model that uses in-context learning to perform complex tasks from human demonstration videos and is designed to operate across multiple robot form factors.

• • Skild AI unveiled S1 as its flagship robot foundation model.
• • The model uses in-context learning to follow a task demonstrated in a single human video.
• • S1 is intended for long-horizon tasks and multiple robot form factors, including quadrupeds, humanoids, and static arms.

Source evidence

Confidence 98% / Completeness 79%

What we captured

Microsoft Research released two open-weight research pathology foundation models designed to make whole-slide analysis and tumor-microenvironment modeling more computationally efficient at population scale.

• • GigaPath-Flash uses a 22-million-parameter ViT-S tile encoder and a 21-million-parameter LongNet slide encoder.
• • GigaTIME-Flash uses the GigaPath-Flash ViT-S encoder for H&E-to-multiplex-immunofluorescence prediction.
• • The models are intended for research and require additional validation before any clinical application.

Source evidence

Confidence 100% / Completeness 91%

What we captured

Perceptron AI released Isaac 0.5, a 36-billion-parameter open-weight embodied foundation model, with model weights, a technical report, and fine-tuning and inference code.

• • Isaac 0.5 has 36 billion parameters.
• • The model was trained on three trillion multimodal tokens, one million hours of general video, and 100,000 hours of robotics experience.
• • Perceptron reported a 97.2% average success rate on the LIBERO benchmark.
• • Model weights, a technical report, and fine-tuning and inference code were released for developers.

Source evidence

What we captured

Intron launched Sahara v2.5 with bilingual code-switching speech recognition across 12 African languages and expanded total language coverage to 31 languages.

• • The update supports bilingual code-switching speech recognition across 12 African languages.
• • Overall speech and voice-agent coverage expanded to 31 languages.
• • Intron reported a 34.3% average word error rate across 12 languages, versus 53.8% for Gemini 3.6.

Source evidence

Confidence 99% / Completeness 79%

What we captured

Amap released ABot-Recon with inference code, checkpoint, and evaluation code. The model uses a 12-frame rolling visual context for long-sequence 3D reconstruction; code is Apache 2.0 and weights are CC BY-NC 4.0.

• • The model uses a rolling 12-frame context for monocular RGB video reconstruction.
• • Amap says it can reconstruct sequences exceeding 10,000 frames.
• • The reported 24.45 FPS benchmark was measured on an H100, not established on the GTX 1080 Ti mentioned in the release.
• • Training code and recipes were scheduled for release on September 30, 2026.

Source evidence

Confidence 99% / Completeness 91%

What we captured

Google launched Gemini 3.5 Transcribe as an API-only speech-to-text model with separate Interactions API and Live API endpoints. The release is in public preview and supports more than 85 languages.

• • The model is split into gemini-3.5-transcribe for recorded files and gemini-3.5-transcribe-live for bidirectional streaming.
• • Reported average WER is 2.6% non-streaming and 4.0% streaming, according to Artificial Analysis.
• • Language detection covers more than 85 languages, including mid-sentence code-switching.
• • Google reports a 70% improvement in time to final transcription versus Chirp 3.
• • The model has no open weights or self-hosted deployment option.

Source evidence