Glossary

Our stack

OrangeLineOur runtime

Our inference runtime, built for agentic workloads and a drop-in alternative to vLLM and SGLang. It manages context during long runs: compressing the KV cache, reusing cached prefixes and suffixes, and freeing GPU memory mid-run. Against standard open model inference, that means 2x faster task completion, a context window past 5M tokens, and 50% to 80% lower cost on long tasks. On your own GPUs it runs 2.3x the concurrent workloads of vLLM at 200k tokens of context. Every model on our API runs on OrangeLine.

Subconscious CacheFeature of OrangeLine

Our KV caching system that matches cached suffixes as well as prefixes. When an agent prunes a span from the middle of its context, a standard prefix cache re-encodes everything after the cut. Subconscious Cache reuses the retained suffix states, so the information the pruned tokens contributed survives implicitly and almost nothing is recomputed.

Auto CompactionFeature of OrangeLine

Compresses the parts of an agent’s context that stopped mattering while the run continues, inside the standard chat completions and messages formats. No manual /compact, no server-side tool calls, no special schemas. Subconscious Cache preserves memory across each compression.

RedLineOur post-training system

Our post-training system, built on top of OrangeLine. We use it to RL-train open models on long, tool-heavy trajectories so they get the most out of runtime capabilities like Auto Compaction and suffix reuse. We also use it to post-train models for customers on their own data and tooling.

Marathon modelsOur model lineup

Open models served on OrangeLine carry the suffix “Marathon.” They keep running faster, cheaper, and more reliably on long agent tasks than the same models on standard inference. GLM-5.3 Marathon and DeepSeek V4.1 Flash Marathon are live on our API today.

GLM-5.3 MarathonLive on our API

The open-source GLM-5.3 frontier coding model, served on OrangeLine and built for agentic coding and long agent runs. Text input. API model string: subconscious/glm-5.3-marathon.

DeepSeek V4.1 Flash MarathonLive on our API

DeepSeek V4.1 Flash served on OrangeLine. Multimodal and built for high-throughput agentic work, with text and image input. API model string: subconscious/deepseek-v4.1-flash-marathon.

Team Token PlanHow to buy

A large monthly token budget with 5x the usage of paying per token, shared across models and across your team. Built for teams pushing a lot of volume, like engineers running coding agents all day. Pay per token is the other self-serve option; enterprise deployments run the full system in your own cloud.

The ecosystem

Qwen

A family of open-weight models from Alibaba, and one of our favorite bases for post-training with RedLine.

Nemotron

NVIDIA’s family of open models, and another base family we like to post-train. Nemotron 3 Ultra is available on our dedicated deployments.

GLM

A family of open-weight models from Zhipu AI (also known as Z.ai), strong on agentic and coding tasks. We serve GLM-5.3 on our API as GLM-5.3 Marathon, and GLM-5.3 Flash on dedicated deployments.

DeepSeek

A family of open-weight models from DeepSeek, known for strong reasoning at low cost. We serve DeepSeek V4.1 Flash on our API as DeepSeek V4.1 Flash Marathon, and DeepSeek V4 Pro on dedicated deployments.

Kimi

A family of open-weight models from Moonshot AI, built for long-context and agentic workloads. Kimi K3 is available on our dedicated deployments.

KV Cache

The key-value tensors a transformer stores for every token in context, so each new token does not recompute attention over the whole history from scratch. It grows with context length, which makes it the memory bottleneck in long agent runs. Most of what OrangeLine does differently starts here.

Inference Runtime

The serving system that actually executes a model on hardware: batching requests, scheduling, managing GPU memory, and caching computed states. vLLM and SGLang are general-purpose runtimes. OrangeLine is a runtime built specifically for agent traffic.

Open Models

Models whose weights are openly released, so you can run them on hardware you control, inspect them, and post-train them. Qwen, GLM, DeepSeek, Kimi, and Nemotron are open model families. Closed models are only reachable through a vendor’s API.

SGLang

A popular open-source inference engine for serving LLMs, known for RadixAttention prefix caching and structured generation. One of the general-purpose systems OrangeLine replaces, and a baseline in our benchmarks.

vLLM

The most widely deployed open-source inference engine, built around PagedAttention for efficient KV cache memory management. The default serving choice on most GPU fleets, and the system OrangeLine most often swaps in for.

llama.cpp

An open-source C/C++ library for running LLMs efficiently on CPUs and consumer or edge hardware, using GGUF quantized weights.