None
None
Long Story Video Skill
Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.
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
Feedback
freeTrialImage.bannerPity
freeTrialImage.upgradeUnlock
- ✓freeTrialImage.benefitHd
- ✓freeTrialImage.benefitWatermark
- ✓freeTrialImage.benefitUnlimited
Inkling 2 Small
Inkling 2 Small pairs native image understanding with a 1M-token window and tunable reasoning effort — test it free in our browser playground.
All Tools
Discover our comprehensive AI-powered animation toolkit
Muse Video
Seedance2.5
The Future of AI Video Is Here.
Seedance 2.0
The Future of AI Video Is Here.
Veo3.1
Create Stunning Videos with Veo3.1
Grok Video Generator
Create Videos from Text or Images with AI
MiniMax H3 video generator
MiniMax H3 AI Video Generator
Seedance 2.0
The Future of AI Video Is Here.
Gemini Omni 1.2 Flash
What Makes Inkling 2 Small Different
Built on a second-generation sparse MoE design, it accepts pictures directly, lets you tune how hard it thinks, and holds a million tokens in view.
- 594B Parameters, Only 17B Fired per TokenInside its 48-layer decoder-only stack, each token is routed to 8 of 384 experts, blending local and global attention layers.
- Tunable Thinking Depth, Million-Token MemorySet the reasoning budget for each call, load up to 1M tokens of material, and supply images sized anywhere from 40px to 4096px per edge.
- Tinker Launch, Apache 2.0 Release to FollowIt arrives on Tinker with weights held back; full checkpoints are slated for an Apache 2.0 drop once outside safety testing and red-teaming wrap up.
Getting Started with Inkling 2 Small
Pick whichever path fits: instant inference, custom fine-tuning, or downloadable checkpoints.
Where Inkling 2 Small Excels
Sustained autonomy, design work that checks its own visuals, and well-calibrated reasoning all carry over into coding, research and knowledge jobs.
Coding Agents That Stay on Task
Reinforcement learning across hundreds of thousands of long coding runs, spanning multiple agent harnesses and compaction strategies, judged on how it works rather than whether tests pass alone.
GPU Kernels and Self-Directed ML Research
Agentic kernel optimization on PMPP-Hard, plus one run that trained Qwen3.5-4B on Tinker for 4.5 hours at $426 to more than double its forecasting score.
Purpose-Built Vision Encoder
A 292M-parameter Perception Encoder stack using 20×20-pixel patches and 2×2 spatial pooling, scoring strongly on OCR, grounding, infographics and visual reasoning.
It Reviews Its Own Renders
The model inspects the output it just produced to refine websites, games and 3D designs — and in 19 of 20 trials it named three wines from a single photo.
Tool Use, Briefs and Finance Benchmarks
MCP tool calls, an ELO of 1527 on AA-Briefcase, and the strongest open-weights numbers across several finance benchmarks covering filings and spreadsheets.
Safety Testing and Dual-Layer Monitoring
CBRN results sit below the open-weight frontier, with a lower Average Risk Score on FORTRESS, backed by a synchronous pre-execution tool monitor plus an asynchronous LLM judge.
Inkling 2 Small: Questions Answered
Quick answers on model size, inputs, training, licensing and customization.
How many parameters does the model have?
17B are active per token out of 594B total, dispatched to 8 of 384 experts across a 48-layer decoder-only transformer.
Is the model open weights yet?
Not at launch — weights stay closed on Tinker, with full checkpoints planned for an Apache 2.0 release the following week once external safety testing and red-teaming finish.
Which input types and context length are supported?
UTF-8 text plus pixel-based images, best between 40px and 4096px per side; output is text, and the context window reaches up to 1M tokens.
What did the training pipeline 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 1,075 joint RL steps and over 17 million rollouts in 9 days.
Where does it rank on benchmarks?
A score of 31 on the Artificial Analysis Intelligence Index, ELO 1527 on AA-Briefcase, leading finance results among open weights, and a top DesignArena Website score still pending.
Can I adapt it to my own dataset?
Yes — it picks up new data faster than earlier models; on a financial-document summarization task it matched Inkling Small's final score in only 75 training steps.
Start Building with Inkling 2 Small Right Now
Jump into the Tinker Playground to run it immediately, spin up fine-tuning from the Cookbook, and keep an eye out for the Apache 2.0 weights.