Inkling 2 Small
AvailableThinking Machines Lab's second-generation multimodal MoE: 594B total / 17B active parameters with text and image inputs (no audio). Scores 31 on Artificial Analysis Intelligence Index v5. Launched Nov 5 2026 hosted on Tinker at $0.30/$1.00 per 1M input/output tokens; open weights followed on Hugging Face on Nov 12 under Apache 2.0, in NVFP4 (native precision, about 500GB of GPU memory) and MXFP8 (about 800GB) checkpoints, with support in SGLang, vLLM, TokenSpeed and Unsloth and third-party APIs including NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. Supports up to 1M-token context and variable reasoning effort.
Specifications
- License
- Open source · Apache-2.0
- Weights
- Downloadable
- Architecture
- Mixture-of-Experts
- Parameters
- 594B · 17B active
- Context window
- 1.0M tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- $0.3 / $1
- Modalities
- TextVisionCode
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.