[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark

What happened

Thinking Machines introduced Inkling 2 Small, a multimodal mixture-of-experts model with 594 billion parameters and 17 billion active at a time. The model was trained on 12 trillion tokens in 68 hours and supports both text and image input, producing text output. It is currently available for use on the Tinker platform, with open weights under the Apache 2.0 license planned for release after external safety testing is completed.

Key details

• Inkling 2 Small has a 1 million token context window and can output up to 1 million tokens.
• The model was pretrained for 68 hours, with the main phase covering 8.4 trillion tokens in NVFP4 precision.
• Pricing on Tinker is $0.30 per 1M input tokens, $1.00 per 1M output tokens, and $1.50 per 1M training tokens.
• Post-training involved 17 million rollouts over 9 days.
• It ranks #7 among open-weight models globally and is #1 among open weights on DesignArena's leaderboard.

Why it matters

Developers and organizations now have access to a large-scale, multimodal model with open weights promised soon, enabling customization and fine-tuning for various applications. The model offers a lower cost per output token compared to some competitors, which may reduce expenses for users. Its performance on several benchmarks and leaderboards positions it as a strong option for those seeking advanced open-weight models outside of China.