Inkling 2 Small Live Workspace
Drop in text or an image and get 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

Run Inkling 2 Small in the Tinker Playground: a multimodal MoE scoring 31 on the Artificial Analysis Intelligence Index, with Apache 2.0 weights due soon.

All Tools

Discover our comprehensive AI-powered animation toolkit

A Closer Look at the Inkling 2 Small Model

This second-generation sparse mixture-of-experts engine reads pictures natively, scales reasoning effort per request and holds a 1M-token window.

  • 594B Parameters, Only 17B Active
    Every token touches just 8 of 384 experts across 48 decoder-only layers, blended with local and global attention for efficiency.
  • Dialable Effort, Million-Token Memory
    Set reasoning effort per request, feed up to a million tokens, and pass images from 40px to 4096px per side for the sharpest output.
  • Tinker-Only at Launch, Apache 2.0 Soon
    The model opens on Tinker behind closed weights; full checkpoints arrive under Apache 2.0 once external safety testing and red-teaming wrap up.

Three Ways to Put Inkling 2 Small to Work

Pick the path that fits: instant inference, fine-tuning, or downloadable checkpoints.

What Inkling 2 Small Can Do for You

Calibrated reasoning, vision-in-the-loop design and long-running autonomy carry across coding, research and knowledge work.

Coding Agents That Keep Going

Reinforcement learning across hundreds of thousands of long-horizon coding tasks and agent harnesses, graded on approach rather than passing tests alone.

GPU Kernels and ML Autoresearch

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

A 292M Vision Stack for Grounding

20×20-pixel patches with 2×2 spatial pooling power strong OCR, grounding, infographic reading and visual reasoning.

It Checks Its Own Renders

The model inspects its own rendered output to iterate on 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 top open-weights results on finance benchmarks covering filings and spreadsheets.

Safety Guardrails and Live Monitoring

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

FAQ

Inkling 2 Small: Your Questions Answered

Straight answers on model size, inputs, training, benchmarks and customization.

1

How big is the Inkling 2 Small model?

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

2

When will the weights go open?

At launch the weights stay closed on Tinker; full weights are planned for an Apache 2.0 release the following week, once external safety testing and red-teaming are complete.

3

Which inputs and context lengths work?

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

4

What did the training run look like?

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 a joint RL run of 1,075 steps and over 17 million rollouts across 9 days.

5

Where does it top the charts?

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

6

Can I tune it on my own data?

Absolutely — it picks up new material faster than earlier models on the same recipe, matching Inkling Small's final score on a financial-document summarization task in only 75 training steps.

Start Building With Inkling 2 Small Today

Launch the Tinker Playground for instant inference, spin up fine-tuning with the Tinker Cookbook, and get ready for the Apache 2.0 weight drop.