AI Model Launch Tracker

What we captured

MiniMax released the model weights publicly on Hugging Face.

• • Weights published October 23, 2026.
• • 428 billion parameters; 23 billion active per token.
• • Repository size is about 252.4 GB.
• • API pricing is $0.15 per million input tokens and $0.60 per million output tokens.

Source evidence

Confidence 99% / Completeness 100%

What we captured

Canva announced its first internally developed image-generation model, available globally through Dream Lab and a public API.

• • Announced October 22, 2026.
• • Canva describes it as its first internally developed image-generation model.
• • Available globally through Dream Lab and a public API.
• • MCP server availability was listed as coming soon.

Source evidence

What we captured

Ai2 released OlmoEarth v1.3, an open satellite embedding model, with precomputed global embeddings and the OlmoEarth Embeddings Explorer.

• • Release announced October 22, 2026.
• • Embeddings are 10-meter resolution and available in 64- and 128-dimensional versions.
• • Ai2 reports the 128-dimensional embeddings ranked highest in 20 of 27 evaluations.

Source evidence

Confidence 99% / Completeness 91%

What we captured

Recursive released Flint, an open-weight mixture-of-experts coding model developed with assistance from its Eureka Machine research system.

• • Recursive calls Flint its first publicly available language model.
• • The model has 40 billion total parameters and 2.5 billion active parameters per token.
• • Weights are available on Hugging Face under Apache 2.0.
• • The article reports Recursive's benchmark claims as not independently verified.

Source evidence

What we captured

Chariot announced a real-time speech recognition model for 10 Indian languages.

• • Chariot announced the model on October 22, 2026.
• • The model is designed to transcribe speech as it is spoken across 10 Indian languages.
• • Chariot reported fewer errors than competing systems in its evaluations.

Source evidence

Confidence 97% / Completeness 79%

What we captured

iFlytek launched the Spark X2.6 end-cloud model and made it available for trial; the article also reports an open-source release of the X2.6-4B on-device model.

• • Launched at the Global 1024 Developer Festival on October 22, 2026.
• • The article reports a 293B-A30B MoE architecture and up to 1 million tokens of context.
• • iFlytek says the model was trained entirely on domestic computing resources.
• • Listed prices: CNY 0.9 per million input tokens and CNY 3.6 per million output tokens.

Source evidence

Confidence 96% / Completeness 96%

What we captured

Recursive unveiled Flint AI on October 21, 2026, and published its weights on Hugging Face under the Apache 2.0 license.

• • The article reports the model was unveiled on October 21, 2026.
• • The weights are available on Hugging Face under Apache 2.0.
• • The article reports 39.8 billion total parameters and about 2.5 billion active per token.
• • Published context figures conflict: 262,144 tokens in the model card and 512,000 in Recursive's post.

Source evidence

Confidence 98% / Completeness 91%

What we captured

Base Labs released two open-weight, Qwen-based decision models designed to return calibrated probabilities rather than generated answers.

• • Released October 21, 2026.
• • Both models are described as open-weight Qwen-based decision models.
• • mammoth-1 has 27 billion parameters; elephant-1 has 9 billion parameters.
• • Base Labs claims mammoth-1 leads the Jev Decision Index.

Source evidence

Confidence 97% / Completeness 79%

What we captured

NVIDIA announced Kumo Relational, a pretrained relational-data model, and said it is available through the structured-data-models library and NVIDIA NIM.

• • Announced October 21, 2026; NVIDIA says available today.
• • Uses up to 10,000 historical examples as in-context evidence and requires no task-specific training.
• • Evaluated across 41 tasks and 15 relational databases, according to NVIDIA.
• • Weights use the OpenMDW License Agreement v1.1; code uses Apache 2.0.

Source evidence

Confidence 100% / Completeness 91%

What we captured

Inception released Mercury 3, a diffusion-based language model, with a context window of up to 1 million tokens and availability through its platform and named third-party services.

• • The article reports that Mercury 3 has been released.
• • Inception reports 1,068 tokens per second on Nvidia GPUs and wins in eight of 12 internal benchmark tests; no independent tests are available yet.
• • Regular prices are $0.20 per million input tokens and $0.70 per million output tokens; launch pricing is half off, with no stated end date.

Source evidence