Learning intelligence
Models that can tell what someone already knows, what they're ready for next, and when they're about to forget it. We treat mastery and memory as things you can measure.
AfrontierAIlabforhumanlearning.
We research and build learning technology at the frontier of AI, so that anyone, anywhere, can learn from the best teacher there has ever been.
Claude Opus 5.5AnthropicClaude Sonnet 5.5AnthropicClaude Haiku 5.5AnthropicClaude Fable 5.1AnthropicGPT-6OpenAIGPT-5.5OpenAIGPT-5OpenAIo3OpenAIo4-miniOpenAIgpt-oss-120bOpenAIgpt-oss-20bOpenAISora 2OpenAIGPT Image 1OpenAIGPT RealtimeOpenAIWhisperOpenAIGemini 4 ProGoogle DeepMindGemini 4 FlashGoogle DeepMindGemini 3 ProGoogle DeepMindGemma 4Google DeepMindVeo 3Google DeepMindImagen 4Google DeepMindNano BananaGoogle DeepMindLyria 2Google DeepMindGenie 3Google DeepMindGrok 5xAIGrok 4xAIGrok Code FastxAIGrok ImaginexAIDeepSeek V4DeepSeekDeepSeek R2DeepSeekDeepSeek V3.2DeepSeekDeepSeek OCRDeepSeekQwen 3.5 MaxAlibaba QwenQwen3-CoderAlibaba QwenQwen3-VLAlibaba QwenQwen3-OmniAlibaba QwenQwen-ImageAlibaba QwenWan 3.0Alibaba QwenKimi K3Moonshot AIKimi K2 ThinkingMoonshot AIGLM-5.3Z.ai (Zhipu)GLM-4.6VZ.ai (Zhipu)CogVideoXZ.ai (Zhipu)Seedance 2.5ByteDance SeedSeedream 4.0ByteDance SeedDoubao Seed 1.6ByteDance SeedSeed-OSSByteDance SeedMiniMax M2MiniMaxHailuo 02MiniMaxMiniMax Speech 02MiniMaxLlama 4 MaverickMetaLlama 4 ScoutMetaSegment Anything 2MetaMistral Large 3Mistral AIMagistralMistral AICodestralMistral AIDevstralMistral AIPixtral LargeMistral AIVoxtralMistral AIHunyuan-T1Tencent HunyuanHunyuanVideoTencent HunyuanHunyuan3D 2.5Tencent HunyuanKling 2.5KuaishouERNIE 5BaiduStep-3StepFunMAI-1Microsoft AIMAI-Voice-1Microsoft AIPhi-4Microsoft AINova PremierAmazonNova SonicAmazonNova ReelAmazonNemotron UltraNVIDIACosmosNVIDIACommand ACohereAya VisionCohereEmbed 4CohereJamba 1.7AI21 LabsGranite 4IBMSonar ProPerplexityOLMo 2Allen Institute for AIMolmoAllen Institute for AIGen-4RunwayAlephRunwayRay3Luma AIFLUX.1 KontextBlack Forest LabsFLUX 1.1 ProBlack Forest LabsStable Diffusion 3.5Stability AIStable Audio 2.5Stability AIMidjourney V7MidjourneyIdeogram 3.0IdeogramEleven v3ElevenLabsScribeElevenLabsEleven MusicElevenLabsAvatar IVHeyGenSuno v5SunoNova-3Deepgramvoyage-3.5Voyage AIjina-embeddings-v4Jina AILFM2Liquid AISolar Pro 2UpstageClaude Opus 5.5AnthropicClaude Sonnet 5.5AnthropicClaude Haiku 5.5AnthropicClaude Fable 5.1AnthropicGPT-6OpenAIGPT-5.5OpenAIGPT-5OpenAIo3OpenAIo4-miniOpenAIgpt-oss-120bOpenAIgpt-oss-20bOpenAISora 2OpenAIGPT Image 1OpenAIGPT RealtimeOpenAIWhisperOpenAIGemini 4 ProGoogle DeepMindGemini 4 FlashGoogle DeepMindGemini 3 ProGoogle DeepMindGemma 4Google DeepMindVeo 3Google DeepMindImagen 4Google DeepMindNano BananaGoogle DeepMindLyria 2Google DeepMindGenie 3Google DeepMindGrok 5xAIGrok 4xAIGrok Code FastxAIGrok ImaginexAIDeepSeek V4DeepSeekDeepSeek R2DeepSeekDeepSeek V3.2DeepSeekDeepSeek OCRDeepSeekQwen 3.5 MaxAlibaba QwenQwen3-CoderAlibaba QwenQwen3-VLAlibaba QwenQwen3-OmniAlibaba QwenQwen-ImageAlibaba QwenWan 3.0Alibaba QwenKimi K3Moonshot AIKimi K2 ThinkingMoonshot AIGLM-5.3Z.ai (Zhipu)GLM-4.6VZ.ai (Zhipu)CogVideoXZ.ai (Zhipu)Seedance 2.5ByteDance SeedSeedream 4.0ByteDance SeedDoubao Seed 1.6ByteDance SeedSeed-OSSByteDance SeedMiniMax M2MiniMaxHailuo 02MiniMaxMiniMax Speech 02MiniMaxLlama 4 MaverickMetaLlama 4 ScoutMetaSegment Anything 2MetaMistral Large 3Mistral AIMagistralMistral AICodestralMistral AIDevstralMistral AIPixtral LargeMistral AIVoxtralMistral AIHunyuan-T1Tencent HunyuanHunyuanVideoTencent HunyuanHunyuan3D 2.5Tencent HunyuanKling 2.5KuaishouERNIE 5BaiduStep-3StepFunMAI-1Microsoft AIMAI-Voice-1Microsoft AIPhi-4Microsoft AINova PremierAmazonNova SonicAmazonNova ReelAmazonNemotron UltraNVIDIACosmosNVIDIACommand ACohereAya VisionCohereEmbed 4CohereJamba 1.7AI21 LabsGranite 4IBMSonar ProPerplexityOLMo 2Allen Institute for AIMolmoAllen Institute for AIGen-4RunwayAlephRunwayRay3Luma AIFLUX.1 KontextBlack Forest LabsFLUX 1.1 ProBlack Forest LabsStable Diffusion 3.5Stability AIStable Audio 2.5Stability AIMidjourney V7MidjourneyIdeogram 3.0IdeogramEleven v3ElevenLabsScribeElevenLabsEleven MusicElevenLabsAvatar IVHeyGenSuno v5SunoNova-3Deepgramvoyage-3.5Voyage AIjina-embeddings-v4Jina AILFM2Liquid AISolar Pro 2Upstage
To build 
intelligence that teaches, and to put it in everyone's hands, for the benefit of society.
A great teacher, for every person who wants one.
Most people never get a great teacher. The ones who do usually remember that teacher's name for the rest of their lives. The gap between the few who are taught well and everyone else is the problem Nucleus Labs was started to work on.
Models can now read, reason, speak, and make video well enough to teach. We take the best of them, test them against what learners actually remember a week later, and build them into tools a school or a company can start using on Monday morning.
We're a small lab based in Vancouver, Canada. We ship products, publish what we learn, and work with every serious model maker, because the learner deserves the best one available.
Push what AI can teach.
Our research sits close to the limit of what models can do. Tutors that reason across a whole course, voice and video instruction in real time, simulations generated on demand, and agents that coach someone through the actual task.
Protect the people it teaches.
Before anything reaches a learner, we check it for accuracy, fairness, privacy, and its effect on the person using it. We write the results down and share them with the people who rely on it.
What we research
Everything we study comes back to one question. How do you help a person really learn something, and know for certain that they have?
Generative teaching
Turning a policy document, a product manual, or a textbook chapter into lessons, practice, voice, and video, while keeping every fact traceable to its source.
Model orchestration
No single model is good at everything. We send reasoning to reasoning models, speech to speech models, and video to video models, and we keep a record of why each choice was made.
Evaluation and safety
We hold learning AI to the standard you'd hold a good teacher to. It should be truthful, patient, fair, and accountable, and we test for each of those before anything ships.
Tutors that talk and act
Voice, avatar, and agent tutors that listen, demonstrate, and coach someone through real work, so one to one instruction stops being a luxury.
Open infrastructure
Datasets, benchmarks, and tools for educational AI that we publish, so schools, nonprofits, and other researchers can build on them instead of starting over.
Nucleus OS
A case study from Nucleus Labs
Studio quality training videos. In minutes, not weeks.
Nucleus OS is the first product to come out of the lab. Give it your links, slides, videos, or documents and it makes finished training from them, then teaches it to your people and shows you what stuck.
app.nucleusinstitute.com
The challenge
A good training video still takes weeks to make.
Slow
Someone writes a script, an agency films it, legal reviews it. A single video can take a month.
Stale
By the time it's finished, the product, the policy, or the pitch has usually moved on.
Stitched
Building, assigning, teaching, and reporting happen in four different tools that don't share data.
And at the end of it, nobody could say whether people had actually learned anything.
The approach
One platform. Context in, finished training videos out.
Underneath, a set of agents built on the lab's orchestration research hands each step to the models best suited to it.
Storyboard
Drop in links, images, videos, documents, or a short prompt. Reasoning models read the material and write a story spine, scene by scene, that you can edit frame by frame.
Editable story spine
Usually handled by
Claude Opus 5.5GPT-6Gemini 4 Pro
1. Hook
2. Context
3. Concept
4. Example
5. Practice
6. Recap
Nothing ships until review checks pass for accuracy, tone, and completeness.
2 minutes
Short overviews
Quick video in your brand for the things that change every month.
- Product launches and updates
- Policy and compliance refreshers
- Sales pitches and onboarding intros
Several modules
Long courses
Full programs with a soundtrack, scene edits, and your brand throughout.
- Onboarding and certification
- Customer academies from your docs
- Leadership programs, live or at your own pace
The outcome
Built lean. Already working.
Minutes
From source material to a finished training video. It used to take weeks.
4 → 1
One login and one dataset in place of an LMS, an LXP, an authoring tool, and a virtual classroom.
48
Integrations with the HR, identity, and sales systems companies already run.
6 mo
From launch to our first design partners signing on.
ReplacesLMSLXPAuthoring toolVirtual classroom
What the lab learned
Orchestration beats any single model.
No model we tested was good at storyboarding, rendering, narrating, and tutoring all at once. Giving each step to a specialist made the biggest difference to quality.
Review checks earn trust.
Checking accuracy, tone, and completeness before anything ships is what made companies comfortable putting their name on generated video.
Draft cheap, finish once.
Keeping 480p drafts separate from the 1080p finish brought the cost of frontier video down far enough for small teams to use it.
What we hold ourselves to
Four rules that decide what we build, which models we use, and what we refuse to ship.
Learning is the only score that counts.
We judge our work by what people can do after they use it. Time on screen and benchmark rankings tell us very little about that.
Use the best model for the job.
We work with every major lab and the open weights community. If another model teaches a concept better than the one we used last month, we switch.
Test it before a learner sees it.
Every system is red teamed for misinformation, manipulation, bias, and harm to young or vulnerable people before it ships.
Build for the learner with the least.
Good instruction shouldn't depend on where you live, what you earn, or which language you speak. We design for the hardest case first.
100 frontier models, one learning layer.
We work with the strongest models from every major lab, including Claude, GPT, Gemini, Grok, Llama, DeepSeek, Qwen, Kimi, GLM, Seedance, and Wan. Closed and open, from North America, Europe, and Asia.
Our orchestration layer scores each one on how well people learn from it, then hands every task to whichever model does it best.
100 models from 38 providers
Anthropic
San Francisco- Claude Opus 5.5Reasoning
- Claude Sonnet 5.5Language
- Claude Haiku 5.5Language
- Claude Fable 5.1Language
OpenAI
San Francisco- GPT-6Reasoning
- GPT-5.5Language
- GPT-5Language
- o3Reasoning
- o4-miniReasoning
- gpt-oss-120bOpenLanguage
- gpt-oss-20bOpenLanguage
- Sora 2Video
- GPT Image 1Image
- GPT RealtimeAudio
- WhisperOpenAudio
Google DeepMind
London · Mountain View- Gemini 4 ProReasoning
- Gemini 4 FlashLanguage
- Gemini 3 ProReasoning
- Gemma 4OpenLanguage
- Veo 3Video
- Imagen 4Image
- Nano BananaImage
- Lyria 2Audio
- Genie 3Video
xAI
Palo Alto- Grok 5Reasoning
- Grok 4Reasoning
- Grok Code FastCode
- Grok ImagineVideo
DeepSeek
Hangzhou- DeepSeek V4OpenLanguage
- DeepSeek R2OpenReasoning
- DeepSeek V3.2OpenLanguage
- DeepSeek OCROpenVision
Alibaba Qwen
Hangzhou- Qwen 3.5 MaxReasoning
- Qwen3-CoderOpenCode
- Qwen3-VLOpenVision
- Qwen3-OmniOpenAudio
- Qwen-ImageOpenImage
- Wan 3.0OpenVideo
Moonshot AI
Beijing- Kimi K3OpenReasoning
- Kimi K2 ThinkingOpenReasoning
Z.ai (Zhipu)
Beijing- GLM-5.3OpenReasoning
- GLM-4.6VOpenVision
- CogVideoXOpenVideo
ByteDance Seed
Beijing · Singapore- Seedance 2.5Video
- Seedream 4.0Image
- Doubao Seed 1.6Reasoning
- Seed-OSSOpenLanguage
MiniMax
Shanghai- MiniMax M2OpenCode
- Hailuo 02Video
- MiniMax Speech 02Audio
Meta
Menlo Park- Llama 4 MaverickOpenLanguage
- Llama 4 ScoutOpenLanguage
- Segment Anything 2OpenVision
Mistral AI
Paris- Mistral Large 3OpenLanguage
- MagistralOpenReasoning
- CodestralCode
- DevstralOpenCode
- Pixtral LargeOpenVision
- VoxtralOpenAudio
Tencent Hunyuan
Shenzhen- Hunyuan-T1Reasoning
- HunyuanVideoOpenVideo
- Hunyuan3D 2.5OpenImage
Kuaishou
Beijing- Kling 2.5Video
Baidu
Beijing- ERNIE 5Language
StepFun
Shanghai- Step-3OpenVision
Microsoft AI
Redmond- MAI-1Language
- MAI-Voice-1Audio
- Phi-4OpenLanguage
Amazon
Seattle- Nova PremierLanguage
- Nova SonicAudio
- Nova ReelVideo
NVIDIA
Santa Clara- Nemotron UltraOpenReasoning
- CosmosOpenVideo
Cohere
Toronto- Command AOpenLanguage
- Aya VisionOpenVision
- Embed 4Embedding
AI21 Labs
Tel Aviv- Jamba 1.7OpenLanguage
IBM
Armonk- Granite 4OpenLanguage
Perplexity
San Francisco- Sonar ProLanguage
Allen Institute for AI
Seattle- OLMo 2OpenLanguage
- MolmoOpenVision
Runway
New York- Gen-4Video
- AlephVideo
Luma AI
San Francisco- Ray3Video
Black Forest Labs
Freiburg- FLUX.1 KontextOpenImage
- FLUX 1.1 ProImage
Stability AI
London- Stable Diffusion 3.5OpenImage
- Stable Audio 2.5Audio
Midjourney
San Francisco- Midjourney V7Image
Ideogram
Toronto- Ideogram 3.0Image
ElevenLabs
London · New York- Eleven v3Audio
- ScribeAudio
- Eleven MusicAudio
HeyGen
Los Angeles- Avatar IVVideo
Suno
Cambridge, MA- Suno v5Audio
Deepgram
San Francisco- Nova-3Audio
Voyage AI
Palo Alto- voyage-3.5Embedding
Jina AI
Berlin- jina-embeddings-v4OpenEmbedding
Liquid AI
Cambridge, MA- LFM2OpenLanguage
Upstage
Seoul- Solar Pro 2Language
Model names and trademarks belong to their owners. This list shows the models we evaluate, integrate, or route to. It does not imply endorsement or partnership.
The most important thing AI will do is help people learn.
If you research, teach, run an institution, or simply want to build this with us, we'd like to hear from you.