🖥️
In-House GPU Cluster
A100-class clusters already running in production; dual vLLM / SGLang stacks, with long-context requests as everyday workloads.
Core Business
Token Farm: self-hosted inference infrastructure
From GPU clusters to disk-level KV cache — we rebuilt the cost structure of long-context inference, so every token costs less to produce.
💾
Cascade Disk-Level KV Cache
Lifts KV cache from GPU memory to NVMe and cluster storage; cache reuse skips repeated prefill and cuts memory footprint by 50–80% (design target).
🔀
Inference Gateway & Scheduling
A unified multi-model gateway: request routing, quota billing, and key & user management — turning cluster capacity into deliverable, metered token supply.
Request intakeGateway routing & quotaKV cache match (hit = compute saved)GPU inference (vLLM / SGLang)Streaming output
20.8×
TTFT Speedup
Public benchmark (cache hit)
96%
Cache Hit Rate
Repeated-prefix reuse
50–80%
GPU Memory Saved
Design-target range
20 GB/s
KV Read/Write Bandwidth
Backed by TB-class cluster throughput; depends on network
* Speedup and hit rate are measured on public benchmarks; KV bandwidth depends on network, backed by TB-class aggregate storage throughput.
Model Serving
We build and run models
From deployment and tuning to production — we operate a full model stack on our own clusters.
🚀
Running Models in Production
DeepSeek V4 Flash and other large models run in production on in-house clusters, with 400K-class server-side context configuration — long context as an everyday workload.
🛠️
Full-Stack Inference Tuning
Dual vLLM / SGLang stacks; KV sharding, quantized deployment, and tiered storage (POSIX / GDS-NvFile) tuned layer by layer for higher memory efficiency and throughput.
🔌
Multi-Model Access & Switching
A unified model abstraction and policy layer: multi-vendor model access, gradual switching, and quota control behind a single API.
Why Now
Inference cost is deciding who wins in AI
Tokens are becoming AI’s most basic commodity — and their price is set by infrastructure.
📈
Inference demand is going exponential
Agents, long documents, retrieval, and multi-turn collaboration dominate real workloads — context length and concurrency are exploding together, and inference compute is the real bottleneck.
💸
Cost is the deciding factor
Model quality is converging; the bottleneck shifts from whether it works to cost per token — and GPU memory is the most expensive link in the chain.
⚡
Cache reuse is the structural answer
Every cache hit is compute saved. Make KV cache shareable, schedulable, and reusable across nodes — and token costs drop structurally.
Products
Two agent product lines, built on our own infrastructure
AI products powered by our inference stack: one for developers, one for enterprises.
🧩Helix Agent
Parallel, observable, long-running AI engineering workflows: three-layer multi-agent orchestration, concurrent SubAgents, and cache + compact context management that keeps long tasks moving.
Parallel agentsMulti-remote workspacesDesktop / Mobile / Web
Explore Helix Code →🏢Helix Flow
AI-native workplace collaboration plus an enterprise agent platform: messaging, tasks, approvals, and an AI assistant woven into daily work — with private deployment, secure connectivity, and centralized governance.
AI-native workplacePrivate deploymentSecurity & governance
Explore Helix Flow →Product Capabilities
What Makes Helix Different
Built for teams actually running engineering workflows — not one-off chat completions.
🧠
Three-Layer Multi-Agent Architecture
A manager agent guards user intent, execution agents drive the work, and SubAgents dive into sub-tasks.
⚡
Concurrent Agent Scheduling
Tool calls and SubAgent tasks run in parallel — multi-module investigations finish markedly faster than serial chat loops.
👁️
Visible Parallel SubAgents
While the main conversation keeps moving, every sub-task reports its state and results. Parallelism is no longer hidden.
📦
Cache + Compact Context Management
Large tool outputs are cached and retrieved on demand, long histories compact automatically — keeping long sessions stable.
🌐
Multi-Remote Workspaces
Local projects, remote hosts, and cloud containers under one roof — workspace-level isolation with a consistent workflow.
🔧
LSP + Tools + Skills, Unified
Code intelligence, terminal, Git, MCP tools, and Skills all flow through one execution loop, from analysis to delivery.
Live Demo
See Helix in Action
Real screenshots from the desktop and mobile apps.
Desktop — Multi-Agent Chat
Curious about the Token Farm, model serving, or partnerships?
Inference infrastructure, model deployment, or agent products — get in touch.