Inkling 2 Mini
AvailableSmallest model in Thinking Machines Lab's second Inkling generation, released Nov 20 2026 with Apache 2.0 open weights on Hugging Face and hosted inference and fine-tuning on Tinker. A natively multimodal (text and image in, text out) looped Mixture-of-Experts transformer: 24 layers whose middle 12 run three times with shared weights (48 layer applications per token), 128 experts with 8 routed per token, hybrid local/global attention. 26.4B total parameters, 3.4B active (about 5.4B effective active with looping), 1M-token context. Trained and shipped natively in NVFP4 with the last layers in MXFP8, a ~20GB checkpoint that runs on a DGX Spark or 32GB+ Apple silicon. Thinking Machines says it leads every open-weights model under 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%) (vendor-reported). A research variable-loop variant was released alongside. Third-party API on OpenRouter via Prime Intellect at about $0.10/$0.40 per 1M tokens; Tinker pricing not stated on the launch page.
Specifications
- License
- Open source · Apache-2.0
- Weights
- Downloadable
- Architecture
- Mixture-of-Experts
- Parameters
- 26.4B · 3.4B active
- Context window
- 1.0M tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- —
- Modalities
- TextVisionCode
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.