Must read
[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark
Inkling 2 Small establishes Thinking Machines as a leading Western open-weights lab while demonstrating extreme training efficiency in NVFP4 and new alignment interventions for reward-hacking during reinforcement learning.
Reported by Latent Space ·
What happened
Thinking Machines has released Inkling 2 Small, a 594B-parameter (17B active) multimodal Mixture-of-Experts model pretrained in 68 hours using native NVFP4 precision on fewer than 12 trillion tokens. Available now on the Tinker platform with Apache 2.0 open weights promised next week after safety testing, the model outperforms Western open-weight peers across coding, finance, and multimodal benchmarks.
Who it’s about
- Thinking Machines
- Inkling 2 Small
- Tinker
- John Schulman
- Soumith Chintala