Specifications
Transcribed from the official model card and the Tinker model catalogue, with the source of each figure noted. Where the two disagree, both numbers appear.
Inkling 2 Small is a 594B-parameter sparse Mixture-of-Experts transformer with 48 layers, routing each token to 8 of 384 experts. It takes text and images as input and returns text, supports MXFP8 and NVFP4 numerics, and the model card advertises a context window of up to 1M tokens — though Tinker currently serves 64K and 256K token variants.
Core specification
- Model name
- Inkling 2 Small
- Provider
- Thinking Machines Lab, Inc.
- Release date
- Nov 5, 2026
- Model type
- Multimodal autoregressive transformer
- Architecture
- 48-layer decoder-only transformer with a sparse Mixture-of-Experts feed-forward backbone. Each token is routed to 8 of 384 experts.
- Attention
- Hybrid local/global attention
- Parameters
- 594B total, 17B active per token
- Vision encoder
- Vision Transformer (ViT), 292M parameters. 20×20-pixel patches with 2×2 spatial pooling, image patches interleaved with text tokens.
- Input modalities
- text, image — image dimensions of 40px to 4096px recommended
- Output modalities
- text
- Context window
- 1M tokens advertised; 64K and 256K served on Tinker
- Numerics
- MXFP8, NVFP4
- License
- Tinker Terms of Use · Apache 2.0 intended around Nov 12, 2026
- Languages
- English and general multilingual support
- Distribution
- Available through Tinker for inference and fine-tuning.
Training
Pretraining ran in four phases. The initial phase covered 8.4 trillion tokens of language, code and image data over 68 hours on NVIDIA GB300 NVL72 racks, natively in NVFP4. A second phase added 3.8 trillion tokens of newer, higher-quality data at 512K context. The third phase mixed in roughly 320 billion tokens of reasoning and agent trajectories at 1M context, and a fourth kept about 34 billion tokens of agent trajectories.
Post-training was a single joint reinforcement-learning run of 1,075 steps and over 17 million rollouts across 9 days — a departure from the first Inkling generation, which used separate reasoning and agentic runs combined by multi-teacher distillation. The task mix spanned real-repository software engineering, terminal work, document creation, STEM question answering, vision, web app building, tool use with simulated users, and safety.
Training data types listed on the model card are text, images and video. Provenance is public internet content, third-party datasets, internally generated data and synthetic augmentation.
Reported context behaviour
| Setting | Tokens | Where |
|---|---|---|
| Advertised maximum | 1M | Model card |
| Served variant | 64K | Tinker catalogue |
| Served variant | 256K | Tinker catalogue |
| Phase 2 training | 512K | Announcement |
| Phase 3 training | 1M | Announcement |
How it places
Full benchmark matrixAcross the 59 benchmark rows where the vendor reports a score, Inkling 2 Small holds the outright highest score in 11 and the second-highest in 18. The wins cluster in finance, legal research and visual design; the losses are mostly in pure software engineering against larger open-weights models.
On Artificial Analysis's independent Intelligence Index — which is not the same figure as the vendor's own AA Index column — the model scores 31 , ranked 4 / 29 in their comparison set. It is notably slow per task at 18.5 minutes against a 7-minute median for comparable models.