Large-scale Pretraining
Data mix, scaling laws, MoE load balance, recipes that stay stable. The aim is more capability from every unit of compute.
Data → Recipe → ScaleM MAMODA · RESEARCH ATLAS / 2026
Mamoda Team
We work on multimodal models, reasoning, generation and editing, agents, and self-evolution—and we take that work into real commercial products. Underneath, we are building a training and inference stack with commercial safety in mind: content understanding, ecosystem quality, and content safety at scale.
01 · Architecture
The Stack
Infrastructure carries the Mamoda models and the agent framework. We put both in production, then keep changing them there.
04In Production
Content Understanding & Ecosystem QualityTagging · Parsing · Quality · Grounding
Explainable DecisionsRejection Reasons · Pinpoint the Violation
One-click FixesPinpoint · Edit Video · Edit Audio
Ad Risk AgentRisk Assessment · Trending Events · Policy Orchestration
Touch GUI AgentIn-app Ad Checks · Game Playtests
03Agentic Framework
Hydra AgentPolicies that write themselves · Dynamic tools
Hydra EvolveAutoPrompt · AutoWorkflow
HarnessAtomic plugins · L0–L2 eval
02Model Foundation
Mamoda Multimodal UnderstandingLong context · Think with text, image, video
Mamoda Video Generation & EditingUnderstand and generate together
Mamoda Omni-modal Understanding & GenerationAudio and video, understanding and generation
Embedding ModelsUniversal multimodal embeddings
01Infrastructure
CPT & SFT TrainingKraken
RL & OPD TrainingBI-veRL
Model InferenceEZReal
Evaluation FrameworkOneEval
Data LoopData Flywheel
Online Continual LearningOnline RL
02 · Scaling Up
Hard Problems
How to take something that works and make it hold—at larger scale, over longer trajectories, and in production.
Long Context & Multimodal Scaling
Efficient attention, context that grows in stages, visual tokens that compress. Text, image, and video in one long-context model.
Long Sequences → Multimodal → MemoryTrustworthy Agentic RL Scaling
Long-horizon environments that you can check, and that are hard to game. Then scale the agent with RL & OPD, skills from several teachers, and a budget on reasoning.
Environment → Rollout → DistillTrain and Serve as One
Parallel training, persistent sandboxes, KV cache, speculative decoding, scheduling. Train, evaluate, and serve as one path.
Train → Evaluate → Serve03 · Research
Research Frontiers
Reasoning, perception, action, evolution. We stay with these frontiers so a new method has to work in the real world—and keep improving once it does.
Publications
Eleven recent papers on multimodal understanding, generation and editing, and agents—ICCV 2025, NeurIPS 2025, ICLR 2026, ICML 2026, ECCV 2026, ACM MM 2026.
FRONTIER 01
Reasoning
Can a model keep improving its judgment—on complex rules, over long context—and improve itself from self-reward, active learning, and live feedback?
Complex RulesLong ContextSelf-RewardActive Learning
Goal: better judgment over time
FRONTIER 02
Perception & Generation
One representation, and one training recipe, for continuous and discrete signals—and models that interleave generation with understanding.
Continuous & DiscreteInterleaved GenerationMultimodal UnderstandingGeneration & Editing
Goal: understand, generate, and edit in one model
FRONTIER 03
Action
How do we give agents better planning, execution, and reflection when they have to act in the real world?
PlanningTool UseLive InteractionReflection
Goal: hard, open-ended tasks
FRONTIER 04
Evolution
How does a model keep learning, evaluating, and improving as the product moves—so the capability compounds?
Self-EvolvingAuto EvalPolicy OptimizationOnline Feedback
Goal: keep up as the world changes