Inkling 2 Small Playground
Send text or an image and read back reasoning, code and finished deliverables
None

None

None

Long Story Video Skill

Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.

Ads Video Skill

Ads Video Skill

Generate professional ads and sales videos—AI auto-generates scripts, prompts, and footage.

3D Science Explainer Video Skill

Convert scientific concepts into stunning 3D explain animations

AI Video Prompt Generator

Feedback

freeTrialImage.bannerPity

freeTrialImage.upgradeUnlock

  • ✓freeTrialImage.benefitHd
  • ✓freeTrialImage.benefitWatermark
  • ✓freeTrialImage.benefitUnlimited

Inkling 2 Small

See how Inkling 2 Small handles image input, step-by-step reasoning and long-context tasks in a free browser playground — no setup needed.

All Tools

Discover our comprehensive AI-powered animation toolkit

What the Inkling 2 Small Model Brings to Your Workflow

Built on a next-gen sparse Mixture-of-Experts design, Inkling 2 Small accepts visual input, scales reasoning effort and works within a 1M-token context window.

  • Sparse MoE Architecture: 594B Total, 17B Active
    A 48-layer decoder-only transformer sends each token to 8 of 384 experts, mixing local and global attention for efficient scale.
  • Flexible Reasoning Effort With a 1M-Token Window
    Adjust effort for each request, process up to 1M tokens and upload images from 40px to 4096px per side for sharper results.
  • Tinker Launch Now, Apache 2.0 Weights Later
    The model debuts on Tinker with closed weights; full checkpoints are planned for Apache 2.0 after external safety testing and red-teaming.

How to Start With Inkling 2 Small

Three practical routes to begin using the model on Tinker.

Core Strengths of the Inkling 2 Small Model

From long-horizon autonomy to vision-in-the-loop design and calibrated reasoning, the model carries across coding, research and knowledge work.

Agentic Coding Across Long Horizons

Reinforcement learning over hundreds of thousands of long-horizon coding tasks, evaluated on approach quality rather than test passes alone.

GPU Kernel and ML Autoresearch

Agentic GPU kernel tuning on PMPP-Hard, plus a 4.5-hour, $426 Tinker run that trained Qwen3.5-4B and more than doubled its forecasting score.

Visual Encoder With Strong Grounding

A 292M-parameter Perception Encoder stack using 20×20-pixel patches and 2×2 spatial pooling, strong on OCR, grounding, infographics and visual reasoning.

Vision-in-the-Loop Output Checks

The model reads its own rendered output to iteratively build websites, games and 3D designs — and named three wines from one photo in 19 of 20 trials.

Knowledge Work and Finance Agents

MCP tool calls, an ELO of 1527 on AA-Briefcase and leading open-weights results on several finance benchmarks for filings and spreadsheets.

Alignment With Two-Layer Monitoring

CBRN scores below the open-weight frontier, a lower FORTRESS Average Risk Score, plus a synchronous pre-execution tool monitor and asynchronous LLM judge.

FAQ

Inkling 2 Small: Common Questions Answered

Quick answers about model size, weights, inputs, training and customization.

1

What is the model's parameter scale?

594B total parameters with 17B active per token, routed to 8 of 384 experts within a 48-layer decoder-only transformer.

2

When will open weights be released?

Weights remain closed on Tinker at launch; full weights are planned for Apache 2.0 the following week after external safety testing and red-teaming.

3

Which input types and context length are supported?

UTF-8 text and pixel-based images, ideally 40px to 4096px per side, with text output and a context window up to 1M tokens.

4

What training process shaped the model?

8.4T pretraining tokens in 68 hours on NVIDIA GB300 NVL72 racks in native NVFP4, a 3.8T-token phase at 512K context, 320B reasoning and agent tokens at 1M context, 34B top-quality agent trajectories, then joint RL for 1,075 steps and over 17 million rollouts in 9 days.

5

Where does it perform best on benchmarks?

A score of 31 on the Artificial Analysis Intelligence Index, ELO 1527 on AA-Briefcase, best-among-open-weights finance results and a pending top DesignArena Website score.

6

Can I adapt it to my own data?

Yes — it adapts faster than prior models using the same data and recipe, reaching Inkling Small's final score on a financial-document summarization task in only 75 training steps.

Start Using Inkling 2 Small in Tinker Playground

Launch the Tinker Playground to run the model instantly, begin fine-tuning with the Tinker Cookbook, and watch for the Apache 2.0 weight release.