Inkling Small
Inkling Small by Thinking Machines: 276B total, 12B active MoE with native text, image, audio, 1M context, relative attention. Open weights, Apache 2.0.
Model Overview
Capabilities, design details, and architectural traits
Inkling Small - Lightweight multimodal MoE with relative attention and native audio-image-text fusion
Inkling Small is a 276B total, 12B active parameter model from Thinking Machines, built as a lighter-weight companion to the full Inkling release. It uses a sparse Mixture-of-Experts backbone where each token routes to 6 of 256 experts plus 2 shared experts active on every token, keeping inference cost low while preserving a large parameter pool.
| Trait | Detail |
|---|---|
| Sparse MoE routing | Each token activates 6 of 256 experts plus 2 always-on shared experts, yielding 12B active of 276B total parameters |
| Relative attention | Replaces RoPE with a learned fourth projection R that encodes key-query distance directly in attention logits |
| Hybrid local-global attention | Decoder layers alternate sliding-window and global attention in a 5:1 pattern |
| Native multimodal fusion | Images enter via a hierarchical patch encoder and audio via discrete token encoding, both projected into a shared hidden space processed jointly by the decoder |
| 1M token context | Supports up to 1 million tokens of mixed text, image, and audio input |
| Multi-format quantization | Ships in BF16, MXFP8, and NVFP4 checkpoints |
| Open-weight customization base | Released under Apache 2.0 with fine-tuning access via the Tinker platform |
The model shares the same architectural recipe as the larger Inkling, including speculative multi-token-prediction layers for faster inference. It is positioned as a cost-efficient open base for developers building agentic systems, coding assistants, and retrieval-augmented applications.
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 30, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 1M | 1M | 1M |
| Pricing (per 1M tokens) | |||
| Input Pricing | $0.30 Best Input Pricing | $4 | $2 |
| Output Pricing | $1.20 Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | textimageaudio | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 25.7 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 52.9 | - | - |
| Agentic Index | 23.5 | - | - |
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
