INN™ Intuitive Neural Network
A brain-closer architecture than Transformer. Three-valued logic, sparse activation, CPU-only inference. Explainability built in, not bolted on.
Traditional deep learning is a black box: you feed it data, it spits out predictions, and you never know why. INN changes that from first principles.
INN adopts a "structural information extraction — logical networking" mechanism, integrating symbolic computation with a data-driven three-valued logic brain-inspired model. It can perform logical reasoning like a human and explain its decision-making process.
This isn't a patch on Transformer. It's a new brain — from neurons to network structure, from activation functions to inference paths, every layer designed for explainability.
Not a bigger
Transformer.
INN is a new brain architecture — every layer redesigned, from neurons to networks, from activation to inference.
Small-sample learning
Extract patterns with minimal data, not millions of labels.
Continuous learning
Learn new knowledge without forgetting, no catastrophic interference.
Multimodal fusion
Process text, image, speech in a unified representation.
CPU inference
No GPU cluster required — runs efficiently on ordinary CPUs.
3-node BIE cluster, CPU-only, no GPU required
Training speed
Inference speed
Training reaches 360,000 tokens/s on a single 600-thread node and 2,160,000 tokens/s across the full 3-node system.
Accuracy
isn't a guess.
Public datasets, reproducible results. Not cherry-picked lab cases — real classification tasks.
Kaggle Diabetes
Medical diagnosis
Kaggle Heart Disease
Disease prediction
Kaggle MNIST
Image recognition
Double Helix
Scientific classification
Six core
advantages.
INN isn't just a faster model — it redefines the baseline across precision, transparency, energy, and generality.
High precision
Consistently outperforms comparable models on standard benchmarks — high accuracy, minimal error.
Model transparency
Every inference step is traceable. The logic chain is visible — no more black box.
Low energy consumption
Native CPU inference. No GPU cluster needed — power draw drops by orders of magnitude.
Fast speed
Full-system inference throughput of 12.6M tokens/s. Response latency under 100ms.
Unified symbolic & numerical
Symbolic logic and numerical computation in a single framework — precision meets generalization.
General-purpose AI architecture
From NLP to image recognition, classification to multimodal — one architecture, many tasks.
Not another
GPU cluster.
Traditional LLM solutions require data centers, GPU clusters, massive power bills and ongoing API costs. INN inside™ packs it all into an appliance under 1m³.
<1m³ device, no GPU required, small-sample training, 6000W standard power, online learning, million-level tokens/s inference, office-friendly private deployment, 3-7.5M CNY investment. vs traditional GPU setups (>1000m³ data center, 200-600M CNY hardware, >30M CNY annual power), INN inside™ fits LLM compute into your office.
Flint
hierarchical architecture.
AI development framework for building and deploying INN models. Generates learning networks based on tasks, determining network nodes, connection structures, and node computation models.
Data Storage Layer
Provides intuitive neural network data storage and metadata management based on underlying CPU, GPU, BPU hardware resources, including KV storage engine, metadata dictionary, neuron node and network structure data access interfaces, and cross-process or multi-machine network communication interfaces.
Network Construction Layer
During task training, responsible for extracting organizational relationships from use case data and performing data annotation fitting. This layer utilizes intuitive neural network perception and cognition components, combined with filtering, classification and other operators, to construct multi-directional, multi-scale neural network graph structures required by tasks based on underlying neuron abstraction interfaces.
Model Services Layer
Provides auxiliary tools and management operation interfaces for the full lifecycle of intuitive neural network training and inference, facilitating rapid construction of various model applications. Mainly includes neural network connection structure visualization tools, data preprocessing and loading tools, and command-line interfaces for training and inference.
Model Applications
As a general artificial intelligence learning algorithm, intuitive neural networks enable the Flint platform to support intelligent application development in various scenarios, such as popular NLP dialogue systems, traditional classification tasks (e.g. Kaggle competition tasks), and multimodal applications combining text and images.
Ease of Use & Performance
Achieving optimal balance between ease of use and performance — developers can get started quickly and use efficiently.
Highly Modular
Each component can be independently replaced or extended, supporting flexible system architecture design.
Rapid Iteration
Supports rapid iterative development with seamless transition from prototype to production.
Intuitive Visualization
Built-in network visualization tools that make complex neural networks understandable.
INN inside™ Basic.
The smallest INN model software version, available for direct download. This is the entry-level version of the INN brain-inspired large model, allowing you to experience the core features of the INN architecture.
Note: INN inside™ PRO and INN inside™ Ultra are only available with BIE PRO and BIE Ultra hardware platforms.
BIE-1:
The Machine.
INN needs a home. BIE-1 Wise One is the purpose-built computing platform for running INN brain-inspired models — hardware, storage, and runtime in one. Mini-fridge footprint. No data center. CPU-only inference.
×BIE-1Compute · Storage · Runtime
Intuitive Neural Network
Brain-inspired · Explainable · CPU-only
Every number,
has meaning.
Every parameter of BIE is carefully designed — from compute power to physical dimensions, from noise levels to energy metrics.
Compute
1,152 cores CPU, 4.8T DDR5 memoryGeneral-purpose X86 architecture — high-throughput inference without dedicated accelerator cards.
Storage
204T hybrid storageMassive local data processing with reduced network transfer latency.
Form factor
Mini refrigerator sizeFits under a desk or in a home corner. No data hall required. Plug and play deployment.
Environment
< 45 dB · < 70°C · ~90% less energyLibrary-quiet operation, excellent thermal control, one-tenth the energy of traditional supercomputing.
Where it
fits,
where it works.
From home to enterprise, healthcare to finance — BIE's compact form factor allows deployment anywhere intelligent computing is needed.
Home Health Management
Personalized health data analysis with explainable-AI health recommendations, not black-box predictions.
Office Work Assistant
Document processing, meeting summaries, knowledge retrieval — runs locally, data never leaves the office.
Medical Diagnosis Assistance
>89% accuracy on scientific classification tasks, fully explainable inference paths to support clinical decisions.
E-commerce Personalization
Local user behavior analysis, real-time recommendation engine — no need to upload user data to the cloud.
Personal Quantitative Trading
Market data analysis and strategy backtesting — brain-inspired algorithms capture non-linear patterns.
Personal Intelligent Partner
An AI assistant that understands you and can explain its own judgments — sitting on your desk.
BIE-1 vs. world-class
supercomputer clusters.
Based on public sources and official data. Comparing M Company RSC and A Company ND H100 v5 Cluster, targeting 600B parameter model workloads.
| Metric | M Company RSC | A Company ND H100 v5 | INN Inside™ + BIE |
|---|---|---|---|
| H100 units / INN units | 24,576 units | 14,336 units | 1 unit (no GPU) |
| Hardware & network CAPEX | $1.0 – 1.2B | $400 – 500M | $500K |
| Infrastructure CAPEX | $500 – 800M | $100 – 200M | 0 |
| Annual energy consumption | 45 – 50 GWh | 26 – 30 GWh | 0.05256 GWh |
| 600B single-node training throughput | 180,000 – 240,000 tokens/s | 105,000 – 140,000 tokens/s | 360,000 tokens/s |
| 600B single-node inference throughput | 3.7 – 4.9M tokens/s | 2.2 – 2.9M tokens/s | 1.35M tokens/s |
Source: public disclosures and official data. Training and inference performance compared on equivalent 600B parameter model compute targets.
One machine,
replacing hundreds of GPU servers.
The following comparison targets equivalent compute throughput (600B parameter model): INN inside BIE-1 vs. an equivalent GPU-H100×8 cluster, full lifecycle cost.
Target: single-node 360,000 tokens/s · 600B parameters
Target: single-node 1,350,000 tokens/s · 600B parameters
Hardware +
software,
complete.
BIE products include both the hardware platform and INN software model. Two configurations for different scale needs.
BIE Ultra + INN inside™ Ultra
- 1152C / 4.8T DDR5
- 204T Storage
- 3 × 24G Graphics Card
- INN inside™ Ultra
- Mini refrigerator size
- < 45 dB noise
- < 70°C at max load
Questions?
Common questions about BIE. If yours isn't here, reach out anytime.
Schedule a
deep dive.
INN offers deployment solutions from single-node to cluster. Fill out the form and our engineering team will schedule a deep-dive session.
What we need
- Deployment scenario (lab / data center / edge)
- Node scale and network topology requirements
- Workload type (SNN training / inference / simulation)
- Whether Flint platform support is needed