Inkling 2 Small
Inkling 2 Small by Thinking Machines: a 594B/17B MoE built for long-horizon agentic coding and autoresearch, with image input, variable effort and 1M context.
Model Overview
Capabilities, design details, and architectural traits
Inkling 2 Small - Thinking Machines' second-generation autoresearch model
Inkling 2 Small opens the second generation of the Inkling family. It is a sparse Mixture-of-Experts model tuned for long-horizon autonomous work: agentic coding, vision-in-the-loop building, and research tasks like autonomously post-training other models.
| Trait | Detail |
|---|---|
| Architecture | 594B total / 17B active MoE; 48-layer decoder-only transformer routing each token to 8 of 384 experts, with hybrid local and global attention layers |
| Native multimodality | Text and image inputs encoded through a vision transformer into a shared hidden space; no audio support, unlike the first Inkling generation |
| Effort and context | Variable thinking effort with a context window of up to 1M tokens |
| Agentic RL training | RL on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses and context-compaction strategies, with an agentic RL grader that judged the model's approach, not just whether code passed tests |
| Autoresearch training | Trained on long-horizon autoresearch tasks including agentic kernel optimization and autonomously post-training models |
| Agent-loop usability | Follows instructions better than the first generation, considers changes carefully, and verifies its work without being asked |
| Availability | Full weights planned for release under Apache 2.0 after external safety testing and red-teaming |
Autoresearch in practice
To demonstrate end-to-end autonomy, Inkling 2 Small was given a time-series forecasting task: autonomously train Qwen3.5-4B. It analyzed the tasks, wrote its own evaluator, experimented with training datasets and recipes, and tried both SFT and RL approaches while debugging along the way. After 4.5 hours it had more than doubled the base model's forecasting score.
Vision-driven building
Its vision and agentic capabilities let it incorporate visual feedback to iteratively build and refine outputs. Demonstrations include a procedurally generated city simulator in Three.js, a from-scratch Python PC emulator that boots a real Linux kernel in the browser, and Blender scenes revised through visual inspection.
Early checkpoints were tested in the real world on OpenCode under the name Fledge-Alpha to validate coding performance inside actual agent workflows.
Benchmark Performance
Independent evaluations · Artificial Analysis
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Thinking Machines | Anthropic | OpenAI |
| Release Date | November 5, 2026 | November 9, 2026 | November 17, 2026 |
| Knowledge Cutoff | - | - | - |
| Context & Limits | |||
| Context Window | 1M | 1M | 1.1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | $0.30 Best Input Pricing | $8 | $8 |
| Output Pricing | $1 Best Output Pricing | $40 | $40 |
| Modalities | |||
| Inputs | textimage | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 30.6 | 64.6 Best Intelligence Index | 62.1 |
| Coding Index | - | - | - |
| Agentic Index | - | - | - |
Benchmark comparison not available
Select at least two models that have available benchmark evaluation data to render the performance comparison charts.
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