Thinking Machines Lab releases Inkling 2 Mini open model
Reported by Hugging Face and The Decoder2 sources ↓
Thinking Machines Lab released Inkling 2 Mini, an open multimodal model with a 1-million-token context window. Its 24-layer architecture repeats 12 layers three times; sources differ on its active parameter count.
Key points
- Inkling 2 Mini repeats 12 of its 24 layers three times
- It accepts text and image inputs and has a 1-million-token context window
- Weights are available on Hugging Face; it can also be tried in Tinker Playground
- It runs locally on Macs with at least 32 GB of memory and Nvidia DGX Spark, The Decoder reports
- The Decoder reports 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2
Sources
Details
The repeated layers are intended to let the model do more internal processing without generating additional text. Thinking Machines says this can help it reach answers using fewer tokens. The model is licensed under Apache 2.0, The Decoder reports.
The Decoder reports that the model scores 48.2% on DeepSWE v1.1 and 46.5% on the Vals Index v2.2, and that Artificial Analysis assigns it 20 points on its Intelligence Index. The model weights and a research version with variable loop counts are available on Hugging Face; it can also be tried in Tinker Playground.
Written from coverage as of
Images
- Hugging Face
- The Decoder
- Hugging Face
- The Decoder