Inkling
Inkling by Thinking Machines: 975B open-weights MoE, 41B active, native text-image-audio input, 1M context, relative attention, and self-fine-tuning on
Model Overview
Capabilities, design details, and architectural traits
Inkling - Open-weights multimodal MoE with relative attention
Inkling is a Mixture-of-Experts transformer, released under Apache 2.0 by Thinking Machines. It is the first model in a planned family, pretrained on 45 trillion tokens of text, images, audio, and video.
| Trait | Detail |
|---|---|
| Relative attention | Uses learned relative position encoding instead of RoPE; each attention layer has a fourth projection producing per-token, per-head relative features modified by key-query distance |
| Hybrid sliding-window attention | Layers alternate in a 5:1 pattern of sliding-window to global attention |
| Sparse MoE with shared experts | 256 experts per layer, 6 routed plus 2 shared experts active on every token; 66-layer decoder |
| Native multimodal encoding | Images via hierarchical patch encoder, audio via discrete token encoding, all projected into a shared hidden space |
| Controllable thinking effort | Balances cost with performance through efficient and adjustable reasoning depth |
| Self-fine-tuning on Tinker | Demonstrated writing, running, and evaluating its own fine-tuning job through the Tinker platform |
| 1M token context | Supports up to 1 million tokens across all input modalities |
Designed for customization
Inkling is explicitly positioned not as the strongest overall model but as an open-weights base optimized for fine-tuning. It ships with BF16, MXFP8, and NVFP4 checkpoints, the latter including speculative multi-token prediction (MTP) layers for faster inference. The model is accessible through Tinker for fine-tuning and through third-party inference providers.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Thinking Machines | Anthropic | Anthropic |
| Release Date | July 15, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 1M | 1M | 1M |
| Pricing (per 1M tokens) | |||
| Input Pricing | $1 Best Input Pricing | $4 | $2 |
| Output Pricing | $4.05 Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | textimageaudio | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 25.0 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 52.1 | - | - |
| Agentic Index | 22.5 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from Thinking Machines
Other models by Thinking Machines
Top AI Models
Leading alternatives by intelligence score
