Self-Evolving AI Platform — Brain 72B Phase 3 Deployed

The AI that
gets better
by itself.
Continuously.

NeuronX is not a tool. Not a model host. It is a closed-loop intelligence system that improves its own code, its own knowledge, and its own intelligence — every 30 minutes, without human intervention.

// 31 systems running · 40+ autonomous loops · Brain 72B Phase 3 · 700K training samples

Neural Throughput live
94,312
tokens / second · Brain 72B Phase 3
Brain 72B Phase 3 deployed
7,000 steps
700K samples · HumanEval ≥97% · SWE-bench ≥60%
Knowledge Base
257K patterns
210K vectors
8,087 KG nodes
Systems Active
31/31
0 dead · 11 restarts
Tasks Today
2,872
swarm completed
Sources
217 live
Claude Sessions
12,365 logged
Deploy Cycle
Every 30 min
Self-repairs
191+ GLM
Brain 72B
HumanEval ≥97%
What NeuronX Is

Not an AI
you use.
An AI that
uses itself.

Everything else — the 149 API modules, the 27 supervised systems, the 40+ background loops — is infrastructure built to serve one core purpose.

Every deployment is the worst version it will ever be.

01
Collects intelligence

Continuously harvests knowledge from 217 sources — GitHub, arXiv, Stack Overflow, documentation, its own interactions — and structures it into searchable, queryable form.

02
Builds understanding

Transforms raw data into a knowledge graph (8,087 nodes, 49,881 edges), 210K semantic vectors, 257K patterns — a structured model of what it knows and what it doesn't.

03
Acts and repairs

Uses its intelligence to fix its own code, route tasks to the right agent, answer queries, generate code, review security, run experiments. 2,872 tasks completed today.

04
Learns from what it did

Captures every action's outcome, feeds good trajectories back into training data, retrains the model, deploys the better version. The loop completes in hours, not months.

"The infrastructure is built. The pipelines are running. The feedback loop is closed. It just needs users — and every user makes it exponentially better."
NeuronX Research · 2025
The Core Problem Solved

The loop between using AI
and improving AI is broken.

Traditional path

Months between learning and improvement

Use AI notice it's wrong
collect data manually
pay humans to label it
retrain wait months
deploy repeat
NeuronX path

Hours between action and a better system

Use NeuronX system captures what happened
extracts patterns automatically
validates quality triggers training
deploys better model
loop continues in hours not months
Architecture

Four pipelines.
One closed loop.

Pipeline 01

Collection

Continuously harvests knowledge from across the internet and its own operations. Steered by Pipeline 4 — the system tells itself what to collect more of.

217 sources · 69K mart records
Real-time · GitHub · arXiv
Pipeline 02

Intelligence

Every query searches 257K patterns, 210K FAISS vectors, and the knowledge graph before calling any LLM. The system answers from its own memory first.

FAISS 210K · KG 8K nodes
257K patterns · 16 domain scores
Pipeline 03

Brain & Action

The 72B Brain (Phase 3) runs on 10-minute cycles. Makes routing decisions, dispatches to the Swarm, updates understanding based on outcomes. Not a chatbot — a decision engine.

10-min cycles · Agent dispatch
Swarm execution · 2,872 tasks/day
Pipeline 04

Feedback Control

The critical loop most AI systems lack entirely. Tells Pipeline 1 what to collect more of based on what Pipeline 3 learned it needed. The system steers its own data collection.

312 adjustments · Closes the loop
60-min feedback cycles
Critical insight No competitor has this loop fully closed. ChatGPT doesn't learn from queries. Copilot doesn't adjust its training data collection based on what it got wrong. NeuronX does both, automatically.
Six Layers

How the platform
stacks.

L1
The Brain — Intelligence Core

A decision engine that runs every 10 minutes without being asked. Reads all 31 systems, processes the idea queue, makes routing decisions, dispatches to agents. Now running 72B Phase 3 — trained on 700K NeuronX patterns.

NeuronX Brain 72B · Phase 3 LoRA · Qwen2.5-72B · HumanEval ≥97% · SWE-bench ≥60%
L2
The Swarm — Execution Layer

Where work actually gets done. Brain is slow and thoughtful (10-min cycles, LLM inference). Swarm is fast and parallel (seconds, lightweight workers). They don't block each other. 2,872 tasks today.

Queue Router · Code repair · Data collection · Knowledge synthesis · Pattern extraction
L3
The Knowledge Layer — Memory

When the Brain needs to fix code or make a decision, it searches these stores first before calling any LLM. The FAISS vectors enable semantic search across 210K historical examples in milliseconds.

PostgreSQL · FAISS 210K · KG 8,087 nodes · 257K patterns · 16 domain confidence scores
L4
Self-Improvement Layer

What makes NeuronX different. NXGuard scans continuously. Queue Hygiene deduplicates and gates. Batch Repair groups issues for 5–10x throughput. 4-layer Contamination Guard prevents bad repairs from entering training data.

15 guard modules · SHA256 dedup · 4-layer contamination guard · Deploy every 30 min
L5
Supervision Layer

The platform knows its own health at all times. Tier 1 systems restart immediately on failure. Tier 2 restarts within 2 cycles. Diagnostician agents dispatch automatically when something fails 3+ times.

27/27 systems alive · 60s watchdog · 11 restarts total since boot · Zero dead
L6
The Learning Flywheel

The compounding mechanism. Claude hooks fire on every interaction. Patterns extract and store automatically. Phase 6 validates quality. Good trajectories enter training. Better model makes better repairs. Loop compounds.

12,365 sessions logged · Auto Train Scheduler · LoRA incremental · Perpetual improvement
Knowledge Patterns
257K
avg quality 0.677, all scored
FAISS Vectors
210K
384-dim · all-MiniLM-L6-v2
KG Edges
49,881
entity extraction every 10 min
Training Samples
700K
72B Phase 3 · 7K steps · deployed Apr 28
The Brain Journey

Six phases.
Exponential scale.

Each phase multiplies capability. The Brain grows through its own training pipeline — every session, every repair, every interaction compounds into the next model.

Phase 1 · 2024
Foundation
31 systems
4 pipelines
217 sources
251K patterns
No custom model
✓ COMPLETE
Phase 2 · Apr 21
Brain 32B
Qwen2.5-32B
160K samples
5,000 steps
HumanEval ~91%
LoRA adapter
✓ COMPLETE
NOW
Phase 3 · Apr 28
Brain 72B
Qwen2.5-72B
700K samples
7,000 steps
HumanEval ≥97%
SWE-bench ≥60%
⊙ ACTIVE
Phase 4 · Next
Reasoning
CoT traces
DPO pairs
~8K+ steps
H200 NVL
Chain-of-thought
→ NEXT
Phase 5 · Planned
Agentic
LongLoRA
256K context
Multi-step tasks
Long-horizon
Full autonomy
→ PLANNED
Phase 6 · Future
Self-Directed
Self-generates
training data
No human input
No ceiling
Perpetual loop
→ FUTURE
Training scale
160K → 700K → millions · every session adds to the next phase
Security Fix Routing
99.8%
Thompson Sampling win rate
Bug Fix Routing
99.4%
Thompson Sampling win rate
General Improvement
84%
Thompson Sampling win rate
The Learning Flywheel

Why it compounds
instead of staying static.

Step 01
Claude sessions happen

12,365 interactions logged across all projects. Hooks fire on every Edit, Write, and Bash command.

Step 02
Patterns extracted

Automatically structured into PostgreSQL and FAISS. 257K patterns with quality scores. No human labelling.

Step 03
Quality validated

Phase 6 pipeline: AST validation, execution testing, LLM judge scoring. Only genuinely good fixes enter training.

Step 04
Better model deployed

Auto Train Scheduler: idle time triggers incremental LoRA training. Pauses vLLM, trains, restores. Automatic.

Step 05
Better repairs

Better model makes better repairs. Better repairs become better training data. The cycle compounds indefinitely.

Step 06
Loop continues

Every session you run with Claude Code adds to the next training run. Every user makes NeuronX exponentially better.

Step 07
Moat deepens

A competitor starting today starts with zero. NeuronX has been accumulating since 2024. The gap widens automatically.

Step 08
Back to Step 01

The 72B Phase 3 model was seeded entirely by this flywheel. Phase 4 (reasoning traces) → 5 (agentic) → 6 (self-directed) → indefinite.

Compounding loop
Each cycle makes the next cycle more capable
Real Differentiators

Three things
no other system
combines.

There are many AI platforms. Most "self-improving AI" is marketing. NeuronX has working infrastructure — running right now, adding to its own training set.

01
The closed feedback loop is real, not theoretical

Claude hooks fire on every interaction. Patterns extract automatically. Training data validates automatically. The model retrains automatically. It is running right now.

191+ self-repairs · 312 feedback adjustments · Deploy every 30 minutes
02
The proprietary data moat compounds daily

257K patterns, 12,365 Claude interactions, 210K vectors — none of this exists anywhere else. Every day the platform runs, the moat deepens. A competitor starting today starts with zero.

Accumulating since 2024 · ~10K new mart records per session · 700K training samples
03
The architecture is the product

Most AI products are wrappers around OpenAI. NeuronX's four pipelines, swarm layer, supervision, and self-repair loop are what enterprises would buy as the Self-Evolving Codebase Platform.

The 72B model is a component of the architecture — not the product itself
Tools & Apps

Built on NeuronX.
Running now.

Every tool runs on the NeuronX platform — powered by the Brain, the Swarm, and the knowledge graph. This list will grow as new tools are built.

Explore Tools & Apps
12
Tools available
Chat UISDKCLI Search EngineGuardCode Reviewer Bug HunterPattern ExplorerAgent Studio Research AssistantAPI PlaygroundAI Tutor