Agent Loop
A reliable, stable, extensible core loop with explicit goals and acceptance criteria.
SOLUTIONS
Enterprise intelligent-R&D solutions
Three paths that land AI inside critical-industry networks: scenario agents, domain model training, and sovereign-stack adaptation — delivered end-to-end on-premise.
01 · AGENT ENGINEERING
Scenario agents: enterprise-grade harness engineering
When an agent lacks task context, how do you supply it fast? When an agent claims a task is done, how do you verify it truly is? Enterprise tasks need more than Q&A — they need a controllable engineering harness that turns raw model capability into verified task completion, replacing "self-declared done" with "verified done".
Context engineering
Necessary-and-precise context per task, with multi-path retrieval and reranking.
Memory management
Cross-session management and persistence of long-term memory.
Sandboxed execution
Local, containerized or remote execution with resource, data and network isolation.
Multi-agent
Focused independent contexts, specialized roles, single-point failure isolation.
Self-verification
Five verification layers — rules, tools, tests, LLM and human — to cancel self-check bias.
Self-evolution
Real task trajectories accumulate; Skills are auto-created and reused; the agent keeps improving.
Safety constraints
Command blocklists, tiered permissions, token budgets and credential hiding, in layers.
Three delivered forms
End-to-end R&D agent
Requirement clarification → design docs → code generation → test execution, fully automated and fused with your processes, tools and domain knowledge.
Software-testing agent
From requirement documents to runnable test scripts, with multi-agent collaboration.
Code-generation agent
From algorithm design documents to runnable engineering code, closed-loop with multi-level test verification and iterative repair.
02 · DOMAIN TRAINING
Domain model training: a model grown from your business
General-purpose LLMs lack enterprise domain knowledge. Domain training teaches the model your code style, business logic and compliance rules.
Your own evaluation set
Collected from your real business code: 90% trains the model, the rest becomes masked real-scenario evals — judged on enterprise data.
Large–small collaboration
Large models are the brain, small models the hands, with smart routing by task complexity; distillation, domain fine-tuning and task-oriented RL train small models that rival large ones in their scenario.
Small base + multi-LoRA architecture
One data pipeline in, multiple LoRA capabilities working together, clean results out.
MEASURED LIFT
Before and after domain training
- Share of code auto-generated10% → 34.93%
- Enterprise-knowledge Q&A accuracy51% → 69%
- Code-generation accuracy20% → 45%
Large–Small Collaboration
Large models think, small models act — complex tasks go to large models, high-frequency tasks to specialized small ones, each doing what it does best.
01
Smart routing
Tasks flow by complexity: completion and change-application go to specialized small models; architecture and cross-module refactors go to large ones.
02
Task atomization
"Implement a feature" decomposes into atomic actions — understand requirements → understand code → design → implement → verify — each matched to its best model.
03
Model specialization
Distillation, domain fine-tuning and task-oriented RL — small models that rival large ones in their own domain.
03 · SOVEREIGN
Sovereign stack: chip to model, under your control
For telecom, aerospace, energy and finance — industries with the hardest requirements on data security and self-reliance — aiXcoder delivers complete enterprise compute matching and assurance: fully private deployment, with models, training and tooling never leaving your network.
Models
Self-developed + open- aiXcoder model family
- DeepSeek
- Qwen
- GLM
Inference engines
Adapted- MindIE
- vLLM-Ascend
- aiXcoder in-house engine
Operating systems
Adapted- Kylin OS
- UOS
- openEuler
Chips
In production- Huawei Ascend 910B
- Hygon DCU
- NVIDIA GPU
Chip certifications & mutual compatibility
- Huawei Ascend
- Hygon DCU
- Iluvatar CoreX
- Enflame
- Moore Threads
- Trust100
- Photoncounts
- PaddlePaddle
Faster inference, MindIE-tuned on Ascend
6×
PagedAttention optimizes memory and compute; FlashDecoding fused kernels + LookAhead parallel decoding cut time-to-first-token from 351.3ms to 96.8ms.
APPLIANCE
The intelligent-R&D appliance
Built on Ascend: R&D productivity metrics, permission management, load balancing, model canary releases and system monitoring — delivered as one box.
- Train & infer in one
- Ready out of the box
- Private deployment
- Security hardened
04 · DEPLOYMENT
How to get aiXcoder
Whether you're a large enterprise or a small team, you start the same way — tell us what you need, and we'll take it from there.
Available now
Private Deployment
Deploy the complete code-AI stack inside your own network — your models and data stay in your hands.
- Enterprise Hub governance + any of the six coding agents + our self-developed code-model family
- Runs on sovereign compute (Huawei Ascend, Hygon DCU) and OS (Kylin, UOS, openEuler); all-in-one appliance available
- Model, data, agents and audit logs all stay on your network — governed and auditable
- Scoped one-on-one to your codebase size, team count, concurrency and customization needs
For teams with hard requirements on data sovereignty and compliance auditing.
On the roadmap
Cloud / SaaS
For smaller teams and individual developers who want a lightweight start.
- Once private deployment is mature, we'll bring the same models and agents to the cloud
- No infrastructure to stand up — ready to use out of the box
- Not open yet — currently in the works
Still on the roadmap. If the cloud edition matters to you, tell us now and we'll notify you the moment it goes live.
05 · ENGAGEMENT
From one conversation to a long-term build.
- 01
Needs assessment
Tell us your team size, stack and goals; we define success metrics together.
- 02
POC on your real code
An evaluation set built from your own codebase, validated at small scale — judged by your engineering metrics.
- 03
Private rollout
A full deployment on your network, integrated with sovereign compute and OS.
- 04
Domain-trained evolution
Models trained on your private data; agents that keep evolving with your workflow.
TRUSTED DEPLOYMENT
Deployed in critical domains
Chip to model, data governance to runtime — every number comes from a real development environment.
Finance
35%+of code AI-generated
Production rollout at a major bank
8M lines written, 100k auto-corrections; 6,000+ users, 2,000+ daily actives — 90% of developers call it genuinely helpful.
Aerospace
20+mission software packages · 9 months
Industry-first control-algorithm software generator
Parses algorithm design documents into hardened, high-assurance C code: 200+ algorithm modules, serving China's lunar missions.
Energy
250k+lines of AI-generated code
China's first nuclear-domain code LLM
15 production projects, 29% of code AI-generated, 80%+ compilation pass rate — fully adapted on domestic chips.
Telecom
70%+test-case acceptance
Software-testing agent at production scale
Requirements → cases → scripts fully auto-generated: 60% execution pass rate, test-script development 50% faster.