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
Test Inkling 2 Small in the browser: native image understanding, tunable reasoning and a 1M-token window, with open weights planned.
All Tools
Discover our comprehensive AI-powered animation toolkit
MiniMax H3
MiniMax H3 AI Video Generator
Seedance 2.5
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
FLUX 3 Video Generator
Kling 3.0
Next-Gen AI Video Generator
Grok Video Generator
Create Videos from Text or Images with AI
MiniMax H3 video generator
MiniMax H3 AI Video Generator
Inside Inkling 2 Small: Sparse Scale With Vision
This next-gen sparse MoE reads images natively, lets you tune reasoning effort, and accepts up to 1M tokens in one pass.
- 594B Total, 17B Active per TokenThe 48-layer decoder-only backbone sends each token to 8 of 384 experts and mixes local with global attention.
- Adjustable Reasoning and 1M-Token InputSet reasoning effort per call, pass up to 1M tokens, and upload images sized from 40px to 4096px on each side.
- Early Access, Then Apache 2.0 WeightsThe model debuts on Tinker behind closed weights; full checkpoints are slated for Apache 2.0 after outside safety review and red-teaming.
Three Paths Into Inkling 2 Small
Pick the route that fits your goal, from instant inference to custom fine-tuning.
What Inkling 2 Small Can Do
From extended agentic coding to visual design loops and careful reasoning, the model carries context across research and knowledge work.
Extended Agentic Coding Runs
Reinforcement learning on vast long-horizon coding work across agent harnesses and compaction methods, scored on process as well as test results.
GPU Kernel and ML Autoresearch
It optimizes GPU kernels on PMPP-Hard and ran a Tinker job that trained Qwen3.5-4B for 4.5 hours at $426, more than doubling its forecasting score.
Native Vision Encoder and Grounding
A 292M-parameter Perception Encoder uses 20×20-pixel patches and 2×2 spatial pooling, performing well on OCR, grounding, infographics and visual reasoning.
Visual Feedback Loops for Output Checks
The model inspects its own rendered pages, games and 3D scenes to refine them, and identified three wines from one photo in 19 of 20 tries.
Knowledge Work and Finance Agents
It makes tool calls through MCP, scores 1527 ELO on AA-Briefcase, and leads open-weights results on finance benchmarks covering filings and spreadsheets.
Alignment Safeguards and Layered Monitoring
CBRN results sit under the open-weight frontier, FORTRESS shows a lower Average Risk Score, and monitoring pairs a synchronous pre-execution tool check with an asynchronous LLM judge.
Inkling 2 Small Questions, Answered
Quick answers on size, inputs, training, benchmarks and customization.
What is the model size and active parameter count?
It has 594B total parameters and activates 17B per token, sending each token to 8 of 384 experts in a 48-layer decoder-only transformer.
When will Apache 2.0 weights be released?
At launch the model runs behind closed weights on Tinker; full checkpoints are planned for Apache 2.0 the next week, after outside safety testing and red-teaming.
Which input types and context length are supported?
It accepts UTF-8 text and pixel images, preferably 40px to 4096px per side, returns text, and supports up to 1M tokens of context.
What did the training pipeline look like?
Pretraining used 8.4T tokens over 68 hours on NVIDIA GB300 NVL72 racks in native NVFP4, followed by a 3.8T-token phase at 512K context, 320B reasoning and agent tokens at 1M context, 34B high-quality agent trajectories, then 1,075 joint RL steps with more than 17 million rollouts in nine days.
Where does it perform best on benchmarks?
It scores 31 on the Artificial Analysis Intelligence Index, reaches 1527 ELO on AA-Briefcase, leads open-weights finance results, and has a pending top DesignArena Website score.
Can I adapt it to my own dataset?
Yes. With the same data and recipe it learns faster than earlier models, matching Inkling Small's final score on financial-document summarization after only 75 training steps.
Start Building With Inkling 2 Small Now
Launch the browser playground to test the model, begin fine-tuning with the Tinker Cookbook, and stay ready for the Apache 2.0 checkpoint drop.