Model Training · Fine-tuning · Deployment
Model Training, Fine-tuning & Deployment
Self-developed model core · distillation · fine-tuning · anti-reverse-engineering--not calling third-party APIs, but自主 training, deployment, and protection.
SELF-DEVELOPED MODELS
Self-Developed Model Matrix
Six core models--covering genomics, proteins, cells, pathways, RNA, and knowledge retrieval
DiVoGenome
pLDDT 92+Genome Triple-Model Cross-Interpretation
Genome triple-model cross-interpretation--fusing variant prediction, functional annotation, and phenotype association, outputting interpretable conclusions at single-nucleotide resolution.
DiVoFold5
pLDDT 95.4Protein Structure Prediction
Protein structure prediction model--self-developed distillation on AlphaFold3 / Protenix architecture, pLDDT 95.4, 3.2x inference throughput.
DiVoCell
10k+ cellsMulti-cell 3D Simulation
Multi-cell 3D simulation model--from single-cell mechanics to tissue-level emergent behavior, supporting 10k+ cell real-time coupling.
DiVoSignal
Graph NNSignaling Pathway Solver
Signaling pathway solver model--encoding pathway graphs as graph neural networks, predicting cascade responses and drug targets under perturbation.
RNALens
Fine-tunedmRNA Structure Fine-tuning
mRNA structure fine-tuning model--fine-tuning secondary structure prediction on RNA-FM backbone, for mRNA vaccine sequence optimization.
Investigate Lens
RAG CoreKnowledge Base Retrieval Engine
Knowledge base retrieval engine--multi-modal literature embedding + hybrid retrieval + citation-level tracing, providing factual foundation for the Agent layer.
TRAINING PIPELINE
Training Pipeline
Six progressive steps--from raw data to encrypted deployment, full-chain self-controlled
Data Preparation
Multi-source data cleaning, annotation, deduplication--gene sequences, protein structures, literature corpora unified into the training data lake.
Pre-training
Foundation pre-training on large-scale unlabeled corpora--learning general representations of sequence, structure, and function.
Fine-tuning
Domain instruction fine-tuning + RLHF--aligning the general foundation to genomics, protein, cell, and other vertical tasks.
Distillation
Teacher-student distillation--compressing large model capabilities to deployable size with controllable loss and multiplied throughput.
Quantization
INT8 / FP8 quantization + sparsification--reducing memory footprint to 1/4 without performance loss.
Deployment
Encrypted inference container deployment on RVDon acceleration clusters--models never leave the trusted environment, inference as a service.
OPEN SOURCE
Model Open Source Strategy
Like the chip, all self-developed models are open source--open source is our core capability, not closure
Fully Open Source Self-developed Models
Like the RVDon chip, all our self-developed models--DiVoGenome, DiVoFold5, DiVoCell, DiVoSignal, RNALens, Investigate Lens--weights and code are fully open source. Open source is our core capability, not closure.
- Weights publicly reproducible
- Training code open
- Apache 2.0 license
Chip-Model Co-Open Source
RVDon RISC-V GPGPU hardware RTL is open source, model inference code is同步 open source--from silicon to weights, the entire acceleration chain is transparent and auditable, community can contribute optimizations.
- RTL open source + ISA public
- Inference code同步 released
- Community can submit PRs
Custom Model On-Demand Release
For B-end client customized domain-adapted models (e.g., specific tissue mRNA optimization, specific cancer type neoantigens), source-code-level delivery per client needs--clients receive complete source code and weights, not a black-box API.
- Source-code-level delivery not API
- Client fully owns customized version
- Can self-deploy and secondary develop
Model Sovereignty Principle
Self-developed models are Gen²AI's core assets. We don't rely on third-party API black-box capabilities--every weight is trained on our own compute, every inference runs within a trusted environment. Model sovereignty means: capabilities are auditable, behavior is explainable, assets are protectable.
CROSS-LAYER
Relationship with Other Layers
The model layer is the pivot of Gen²AI's four-layer architecture--hardware supports it below, Agents wrap it above
RVDon PF extension instructions accelerate triangle matrix multiplication and Flash Attention at the silicon level, model training and inference run directly on self-developed accelerators.
View L1 Engine LayerModel layer capabilities are wrapped by the Agent layer as callable tools--DiVoFold5, Investigate Lens all serve as Agent backend capabilities.
View L3 Agent LayerAgent orchestration results are delivered to end users through the computing service layer--from inference to report, end-to-end on own compute.
View L4 Computing ServicesL1 engine layer provides RVDon acceleration hardware -> L2 model layer trains and infers on it -> L3 Agent layer wraps models as callable tools -> L4 computing service layer delivers results to users. The model layer is the capability hub--hardware accelerates it, Agents wrap it, computing distributes it.
USER SCENE
Who These Models Ultimately Serve
B2B2C: Each model, through dry-computing pipelines, ultimately reaches these end-user scenarios
Build Your Capabilities with Self-Developed Models
No longer relying on third-party black-box APIs--from training to deployment, model sovereignty is in your hands.