AI Model Launch Tracker

What we captured

LG AI Research made hosted API access to K-EXAONE 2.5 available for free through the EXAONE API Platform.

• • LG AI Research announced free API access on November 9, 2026.
• • The model is available through an OpenAI-compatible chat completions endpoint.
• • The documentation lists 31 billion active parameters and Korean and English support.

Source evidence

Confidence 91% / Completeness 87%

What we captured

Fast Simon announced its Commerce-Native AI model for eCommerce product discovery, available immediately and aimed at Shopify brands.

• • The model is for eCommerce product discovery and targets brands using Shopify.
• • It uses shopper behavior, catalogue data and commerce conditions to optimize rankings and recommendations.
• • Merchants can define guardrails for its decisions.
• • Fast Simon said the model is available immediately.

Source evidence

Confidence 94% / Completeness 87%

What we captured

Video Rebirth shipped HyperFlow H3 Flash as open weights, describing it as a 14B-parameter compression of MiniMax-H3 for short-step video generation with synchronized stereo audio.

• • The article dates the open-weights release to November 6, 2026.
• • The model is described as a 14B-parameter compression of MiniMax-H3.
• • The article reports four-to-eight-step video generation with synchronized stereo audio.

Source evidence

Confidence 95% / Completeness 91%

What we captured

Wizstar announced its in-house reference-to-video model, Wizshot 1.0, and made it available to logged-in users.

• • Wizstar described Wizshot 1.0 as its first in-house video generation model.
• • The model supports up to nine reference images and 720p portrait or landscape video generation.
• • Wizstar lists generation at 5 Credits per second and claims a 72% lower cost than Seedance 2.0.

Source evidence

Confidence 99% / Completeness 87%

What we captured

Kakao unveiled its multimodal decision model Kanana-2-v-decide and made it available to download from Hugging Face under a license permitting commercial use.

• • Unveiled November 6, 2026.
• • Analyzes text and images for AI-agent decision tasks.
• • Kakao reported an average response time under 0.04 seconds in a single-Nvidia-H200 test.
• • Available to download from Hugging Face under a Kakao license permitting commercial use.

Source evidence

Confidence 99% / Completeness 91%

What we captured

Inkling 2 Small became available for inference and customization through Tinker; weights remain closed pending safety evaluations, with an Apache 2.0 release planned.

• • The article dates the introduction to November 5, 2026.
• • The model has 594 billion total parameters and activates 17 billion per token, as stated by the company.
• • The article reports an Artificial Analysis Intelligence Index score of 31 and a Surge AI EnterpriseBench-CoreCraft score of 71.8%.
• • Weights were not downloadable at launch; the company targeted an Apache 2.0 release for the following week, subject to safety evaluations.

Source evidence

What we captured

OpenAI released GPT-6 Cyber, a cybersecurity model initially available to selected customers approved through Daybreak Red.

• • OpenAI announced the release on November 5, 2026.
• • The model initially reaches selected Daybreak Red customers.
• • Published standard API prices are $20 per million input tokens and $100 per million output tokens.

Source evidence

What we captured

Nace AI released Drex Max, a 27-billion-parameter decision model, and made the Drex family available through GitHub and Hugging Face under the Open RAIL-M license.

• • The article says the report and model were released around November 5, 2026.
• • Drex Max has 27 billion parameters.
• • Nace AI reports a 67.52 score on Decision Index 0.4 and median latency of approximately 183 milliseconds.
• • The Drex family includes a 9-billion-parameter variant and is distributed through GitHub and Hugging Face under the Open RAIL-M license.

Source evidence

Confidence 98% / Completeness 95%

What we captured

Underdog AI demonstrated Woof, a 31-million-parameter, 31.7 MB text-to-speech model, and says it is available to download and can run locally on most devices.

• • X post showing the demo is dated November 5, 2026.
• • The article describes Woof as 31 million parameters and 31.7 MB.
• • Underdog AI says the model can run on most devices without an internet connection.
• • The article says the model is available to download; no license is specified.

Source evidence

Confidence 91% / Completeness 87%

What we captured

Baidu published the 280B-parameter pre-training checkpoint, with 14B active parameters per token and custom commercial terms.

• • 280B total parameters; 14B active parameters per token.
• • Baidu claims 1M-token inference and says the model was pretrained on 33.5T tokens.
• • Custom terms exclude the EU, UK, and South Korea and require written authorization above $50M in related commercial revenue.

Source evidence

Confidence 97% / Completeness 95%