Sovereign Models
Expertise through engineering

On-device parameter capacity is limited, so expertise comes from engineering: Day0+ adaptation to leading foundation models enables access on release day; task-specific small models are fine-tuned for roles; speculative decoding accelerates local inference; and scenario selection determines the model and execution location for each task. Models run on enterprise-owned devices, accumulating expertise as enterprise assets.

Model adaptation & Optimization

Day0+
Rapid multi-model adaptationSupport Qwen, DeepSeek, GLM, and other leading foundations without relying on one model; integrate updates promptly into professional workflows.
SFT
Domain fine-tuning & Efficient LoRA adaptationFine-tune with specialized legal, tax, and accounting corpora and expert annotations to create role-specific small models. LoRA (Low-Rank Adaptation) reduces training compute costs.
4/8bit
Quantized deploymentCompress models to 4-bit or 8-bit precision for professional inference within device compute and memory budgets.
SPEC
Speculative decoding accelerationSupport MTP, DFlash, and DSpark: draft mechanisms propose multiple tokens at once, and the main model verifies them in parallel, preserving outputs while multiplying speed.
RAG
Private-data augmentation & Long-term memoryClean and integrate private corpora through RAG (Retrieval-Augmented Generation), connect business systems through MCP (Model Context Protocol), and retain project context as long-term memory.

Scenario-based selection

Typical tasks	Recommended models	Execution location	Selection rationale
Document recognition & StructuringInvoices, contracts, and case files	Specialized small models + Dedicated OCR acceleration	On device	Frequent, low-latency, sensitive data
Compliance risk & Indicator matchingTax risk, supply chain compliance	Domain models + Dynamic indicator libraries	Device · Edge	All data stays in the domain, with traceable results
Professional Q&A & Document draftingEveryday legal, tax, and accounting work	LoLR Domain Model Logos LoLR + Speculative decoding	Device first	Balance expertise and response speed
Complex analysis & Cross-domain reasoningM&A due diligence, cross-border structures	Large foundation models + Multi-agent collaboration	Organization-wide compute · Authorized cloud	Model size, concurrency, and cost

Four selection criteria

Data sensitivity

Latency requirements

Model size

Cost

Support for three speculative decoding approaches

On-device decoding is limited by memory bandwidth: generating each token requires reading all weights. Speculative decoding uses a lightweight draft mechanism to propose multiple tokens, then verifies them in one parallel pass through the main model, retaining only consistent candidates. Outputs remain unchanged while speed can increase severalfold. The LoLR model system supports three mainstream approaches.

MTPMulti-Token PredictionMulti-Token Prediction

Prediction heads trained jointly with the main model draft upcoming tokens from the current hidden state. When built into the weights, they add almost no GPU memory overhead.

Best for: foundations with built-in MTP heads · Memory-constrained devices

DFlashBlock-diffusion draftsBlock Diffusion Drafting

A lightweight block-diffusion draft model uses the main model's contextual features to generate an entire draft block in one parallel forward pass. Draft models can be trained separately for different target models.

Best for: multiple foundations and inference frameworks · Long-form generation

DSparkConfidence-scheduled speculative decodingConfidence-Scheduled · Semi-Autoregressive

A semi-autoregressive draft model preserves token dependencies within blocks. Verification length adjusts dynamically to each request's confidence, terminating low-quality drafts early.

Best for: highly concurrent services · Department and institution nodes

All three provide lossless acceleration, with final output determined by the main model's verification;LoLR.AI OS selects automatically based on the foundation model, deployment framework, and node concurrency.

Sovereign Models

Foundation adaptation

Specialized fine-tuning

Inference acceleration

Scenario-based selection

Private operation

Corpora, fine-tuning, acceleration, model selection, and execution location together determine output usability.

Explore LoLR Large Model Logos

LoLR →

Within your enterprise's own boundaries
Deploy this architecture

Tell us your task types, data volume, and data boundaries. Our FDE team starts with an on-site assessment to define your device, model, and Harness solution together.

Explore industry solutions

LoLR HPCs · Trusted Foundation

A complete on-device computing lineup from portable devices to organization-wide servers.

LoLR Large Model Logos LoLR · Sovereign Models

Domain expertise, foundation-model adaptation, device–cloud collaboration, and distributed scheduling.

LoLR Operating System LoLR.AI OS · Enterprise Harness

Agent factory and Harness delivery tools: create, train, manage, and use experts.

LoLR Mobile LoLR.AI APP

Mobile work, agent deployment, and on-the-go access bring the system into the field.

Industry solutions

Explore scenarios, solution components, and devices by industry and role.

Service & Support