AI Model Launch Tracker

What we captured

Pika announced Pika H3 Flex, a post-trained MiniMax H3 model, and released its generator adapters and inference code as open weights. It is available through Pika API Club.

• • Pika describes Flex as a post-trained MiniMax H3 model.
• • Generator adapters and inference code are released as open weights.
• • The model supports text-to-video, first- and last-frame generation, and multimodal reference-to-video at 480p and 768p.

Source evidence

Confidence 100% / Completeness 87%

What we captured

Mistral AI released Large 4 publicly, with weights available on Hugging Face under a modified MIT-style license.

• • Released to the public on October 29, 2026.
• • Weights are available on Hugging Face under a modified MIT-style license.
• • Approximately 1 trillion total parameters, with 52 billion active per token.
• • Supports a 1-million-token context window.
• • API pricing is approximately $0.70 per million input tokens and $2.10 per million output tokens.

Source evidence

What we captured

Baseten released joey-1, a 2B-parameter decision model developed by Base Labs, through Baseten Model APIs and as MIT-licensed open weights.

• • Released October 29, 2026; 2B parameters.
• • API pricing is $0.01 per million input tokens; cached input and output are listed at $0.00.
• • Model card reports a 63.30 score and 37th place among 230 models on the Jev Decision Index 0.4 public split.

Source evidence

Confidence 99% / Completeness 100%

What we captured

Mistral says the finalized open-weight Le Chonk model is freely available to anyone.

• • Mistral says the finalized version is freely available.
• • Mistral describes Le Chonk as a trillion-parameter model.
• • Mistral claims it matches or outperforms models from DeepSeek and Z.ai in coding and cybersecurity.

Source evidence

Confidence 95% / Completeness 87%

What we captured

Apex Intelligence released its first model, Apex-35B-V1.0, a 35B general-purpose model developed with automated AI research. Weights are available on Hugging Face.

• • The company identifies Apex-35B-V1.0 as its first model.
• • The model is built on Qwen3.6-35B-A3B and further developed through mid-training and post-training.
• • The article reports model weights on Hugging Face and code and deployment materials on GitHub.
• • The article does not specify the model's license.

Source evidence

What we captured

Articul8 launched Bivia, a multimodal decision model for enterprise AI agents, available via API.

• • The article reports that Articul8 launched Bivia.
• • The article says Bivia is available through an API.
• • The model supports five decision types and multiple data formats.

Source evidence

Confidence 99% / Completeness 87%

What we captured

Rime announced Prism, a fine-tunable enterprise speech-to-speech model that it says can run on-premises or in a private cloud and be trained on customer calls.

• • Rime announced Prism on October 29, 2026.
• • Rime describes Prism as fine-tunable and says it can be deployed on-premises or in a private cloud.
• • Rime reported Revmo trial results comparing Prism with OpenAI's GPT.

Source evidence

What we captured

Gnani AI launched its speech-to-text model Prisma v3.0, available through its developer APIs.

• • Supports nine Indian languages and Indian English.
• • Provides Verbatim, Transcribe, Code-mix, and Transliterate modes.
• • Gnani AI reported a 16.9% macro word error rate on nine Indian languages under the Vistaar scoring protocol.

Source evidence

Confidence 99% / Completeness 87%

What we captured

H2O.ai released TimeH2O-v1, a time-series forecasting model available under a non-commercial license for internal benchmarking and testing.

• • H2O.ai said TimeH2O-v1 ranked first among 32 models on fev-bench's default leaderboard view.
• • H2O.ai reported a 91.9% win rate and a skill score of 50.8 across 100 forecasting tasks.
• • H2O.ai said the model has 108 million parameters.
• • The model is available under a non-commercial license for internal benchmarking and testing.

Source evidence

Confidence 99% / Completeness 91%

What we captured

Google announced Gemini 4 Carbon on October 28, 2026, with initial access for selected cyber defenders through its Fairwind program. Wider access was planned, but the model was not yet available to developers.

• • Google announcement date: October 28, 2026.
• • Initial rollout targeted a subset of cyber defenders through Fairwind.
• • Artificial Analysis reported an Intelligence Index V5 score of 59 and an AA-Omniscience score of 54.
• • Reported standard pricing: $4 input and $20 output per million tokens.

Source evidence