Model Release·Open Weights·Multimodal
Inkling-Small: Thinking Machines' Efficient Open Multimodal MoE
Inkling-Small is an open multimodal MoE (276B/12B active, Apache 2.0) with up to 1M context—available on Tinker for fine-tune and playground chat.
At a Glance
- Provider: Thinking Machines Lab, Inc.
- Params: 276B total, 12B active (6 of 256 experts + 2 shared)
- License: Apache 2.0
- Modalities: Text, image, audio in; text out
- Context: Up to 1M tokens
- Numerics: BF16, MXFP8, NVFP4
- Hardware (vendor): BF16 ≥ ~600 GB agg. VRAM (e.g. 8× H200 / 4× B300); NVFP4 ≥ ~180 GB
- HF:
thinkingmachines/Inkling-Small,thinkingmachines/Inkling-Small-NVFP4 - Distribution: Hugging Face + Tinker API / Playground
Selected scores (vendor table)
| Signal | Inkling-Small | Note |
|---|---|---|
| SWE-Bench Verified | 80.2% | vs Inkling 77.6%; bash-only harness (vendor footnote) |
| Terminal-Bench 2.1 | 64.7% | Internal harness (vendor footnote) |
| HLE text-only | 31.6% | Ahead of Inkling 29.7% on vendor table |
| SimpleQA Verified | 20.6% | Trails Inkling 43.9% (knowledge/factuality gap) |
Label self-reported / harness-specific rows. Independent LLM Stats verification pending.
What's New vs Inkling
- Same multimodal recipe at ~12B active vs 41B.
- Later training recipe + on-policy distillation from Inkling; vendor says it beats Inkling on some reasoning/agentic benches while Inkling keeps the knowledge lead.
- Day-one Tinker fine-tune + playground (text/image/audio).
When to Use It
Good fit: Open multimodal agents where Apache 2.0 and fine-tuning matter; cost-sensitive coding/tool loops vs full Inkling.
Not automatic: If you need Inkling's factuality/knowledge coverage, keep the larger sibling (or a closed model) in the mix.
Sources
- https://thinkingmachines.ai/news/inkling-small/
- https://thinkingmachines.ai/model-card/inkling-small/
- https://huggingface.co/thinkingmachines/Inkling-Small
- https://huggingface.co/thinkingmachines/Inkling-Small-NVFP4
Questions
Frequently Asked Questions
- July 30, 2026.
- 276B total parameters with about 12B active (6 of 256 experts + 2 shared).
- Apache 2.0.
- Text, image, and audio in; text out. Context up to 1M tokens.
- Hugging Face weights plus Thinking Machines' Tinker API / Playground for fine-tune and chat.
- Vendor tables show roughly Inkling-class agentic/reasoning at ~12B active vs Inkling's 41B, with Inkling still leading on knowledge/factuality (e.g. SimpleQA). Label harness-specific rows carefully.
Continue Reading