# Subconscious: Inference Systems Designed for Agents > Built for agents, not chatbots. Run open models on infrastructure you own, in your cloud or on the edge, with faster throughput, more concurrent workloads, and longer-context reasoning on the same hardware. Subconscious is an AI lab and infrastructure company born from research at MIT CSAIL. We build OrangeLine, an inference runtime purpose-built for agentic workloads, and Redline, a post-training system co-designed with it. OrangeLine prunes and compacts the KV cache on the GPU during a run. Where general-purpose engines stop to compact when they hit the context limit, OrangeLine compacts at runtime and keeps going, so a node does the work that used to take two. We install the runtime inside your cloud so your data never leaves, license it on edge devices, and also serve open models on a hosted, OpenAI- and Anthropic-compatible API so you can trial before you deploy. Based in Cambridge, Massachusetts. Spun out of MIT CSAIL. - Website: https://www.subconscious.dev - API Docs: https://docs.subconscious.dev - Hosted API: https://api.subconscious.dev - Drop-in replacement for vLLM and SGLang. --- ## Key Numbers - **3.5x** faster token throughput than vLLM at maximum context length, on the same models and hardware - **2.3x** the concurrent workloads on the same compute footprint (roughly half the GPUs for the same work) - **10x** the stateful context window of the base model, via runtime KV-cache compression - **50%+** lower latency per task via the Subconscious Cache (prefix and suffix caching) - **95%+** cache-hit rate observed on agentic coding workloads, so most input tokens bill at the cached rate --- ## How It Works ### OrangeLine: the inference runtime OrangeLine is an inference system optimized for agents rather than chatbots. It treats long agent traces as the primary workload: - **Runtime context compression**: prunes and compacts the KV cache on the GPU as the agent works. No stop-the-world compaction, no manual summarization, no context-window overflow. - **Subconscious Cache**: prunes context mid-run while preserving suffix states, so information is reused without re-encoding. Prefix and suffix caching raise cache-hit rates and keep agent cost curves nearly linear. - **Drop-in**: a replacement for vLLM and SGLang. Same models, same hardware, same OpenAI- and Anthropic-compatible API surface. - **Stateful context window** extended to roughly 10x the base model's limit, so agents hold far more context and reason deeper on long-horizon tasks. ### Redline: model post-training Redline is a post-training system co-designed with the OrangeLine runtime, so models learn to exploit runtime compression and suffix reuse: - Trains on your data and tooling, producing a model unique to your workload. - Extremely efficient, using a fraction of the usual compute. - Long-horizon by design: models learn from full reasoning traces, not single shots. - Supports SFT, OPD, and RL on agentic workloads, with hands-on support from our research team. The TIM (Thread Inference Model) family, such as TIM-Qwen3.6 27B, comes out of this approach: models post-trained for recursive, decompositional reasoning and served on OrangeLine. --- ## Who It's For ### Inference clusters / GPU fleets: https://www.subconscious.dev/enterprise "Your GPU fleet, doing several times the work." OrangeLine is a drop-in replacement for vLLM or SGLang on GPUs you control: same models, same hardware, several times the output (3.5x throughput, 2.3x concurrent workloads, 10x stateful context). Priced per GPU, any GPU type, any fleet size (1 to 10,000 GPUs), installed and supported by our team. ### Coding agents: https://www.subconscious.dev/coding-agents "Power coding agents, 100% in your cloud." We run the open-source frontier model GLM-5.2 on the runtime, so engineers use a top-tier coding model with the benefits of an on-prem deployment, at less than half the GPUs of other inference systems. Privacy (code and prompts never train other models), no refusals or silent degradation, and uptime you control (no third-party rate limits). Works with Claude Code, Codex, Cursor, OpenCode, Pi, Zed, and any coding harness or inference SDK. ### Product agents: https://www.subconscious.dev/product-agents "Build product agents that turn a profit." Run enhanced open models in your cloud, so products are capable enough for real users and cheap enough to make money when they use it. Good for data-heavy agents (research, coding, document understanding), multi-modal agents (browser automation, computer use, image generation), workflow automation across long MCP and tool chains, and latency-sensitive use (translation, voice agents). OpenAI- and Anthropic-compatible, on a dedicated endpoint that scales with traffic. --- ## Available Models Two models are live on the hosted API today; more open models run on the runtime for cloud deployments. | Model | API name | Pricing (per 1M tokens) | Notes | |-------|----------|-------------------------|-------| | **TIM-Qwen3.6 27B** | `subconscious/tim-qwen3.6-27b` | $0.15 cached / $0.30 input / $3.00 output | Post-trained on Qwen3.6 27B, served on OrangeLine. Efficient and capable for agentic workflows. | | **GLM-5.2** | `subconscious/glm-5.2` | $0.26 cached / $1.40 input / $4.40 output | Open-source frontier coding model, served on OrangeLine, built for agentic coding and long-running agent work. | Beyond the hosted API, the runtime serves most open models (Qwen, Nemotron, Gemma, DeepSeek, GLM) for in-cloud deployments, or bring your own model to post-train with Redline. Cached input tokens are billed at a steep discount to the input rate. With efficient caching we see upwards of a 95% cache-hit rate on agentic coding workloads, so most input tokens bill at the cached rate. --- ## Pricing Subconscious's primary offering is running open models on infrastructure you own; the hosted API is a way to trial first. ### In your cloud Half the GPUs, faster throughput, better capability, better economics. We install the runtime inside your VPC so a node does the work that used to take two. You keep the savings, we charge a fraction of it, with hands-on setup by our research team. Priced per GPU node. Any open model. If you have GPUs we install in your cloud (typically within a week); if you need GPUs we source hardware via partners first, then install. ### Hosted API (trial) Run open models on our hosted, OpenAI- and Anthropic-compatible API before you deploy. $50 in free credits on signup. No credit card required. Pay-as-you-go per-token pricing (see Available Models). --- ## API Reference The Subconscious hosted API is **OpenAI- and Anthropic-compatible** and available at `https://api.subconscious.dev`. ### Authentication ``` Authorization: Bearer YOUR_API_KEY ``` Get your API key at https://www.subconscious.dev. $50 in free credits on signup. No credit card required. ### Quick Start: Python (OpenAI SDK) ```python from openai import OpenAI client = OpenAI( base_url="https://api.subconscious.dev/v1", api_key="YOUR_API_KEY", ) response = client.chat.completions.create( model="subconscious/tim-qwen3.6-27b", messages=[ {"role": "user", "content": "Explain how transformer attention works."} ], ) print(response.choices[0].message.content) ``` ### Quick Start: cURL ```bash curl -X POST https://api.subconscious.dev/v1/chat/completions \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "subconscious/tim-qwen3.6-27b", "messages": [ {"role": "user", "content": "Explain how transformer attention works."} ] }' ``` Full documentation: https://docs.subconscious.dev --- ## Platform Features ### Playground Browser-based chat interface for testing prompts and tools: streaming responses with full reasoning traces, MCP server connections, and a toggle for thinking/reasoning mode. ### API Key Management Create, rotate, and revoke API keys per organization. ### Billing & Usage Credit-based billing in microdollars, real-time usage dashboards with per-model breakdowns, auto-pay with configurable thresholds, and automatic key revocation when a balance goes too negative. ### Tools - **Hosted tools** (built-in, no setup): web search, web scrape, deep research, code sandbox. - **Authenticated integrations** with managed OAuth: Gmail, Slack, HubSpot, Linear, Notion, Google Drive, and more. - **Custom tools**: any HTTP/REST endpoint, any MCP server, or import from an OpenAPI spec. --- ## Core Research We research both sides of the stack, the model and the runtime, on the thesis that the runtime layer is under-innovated. **"Beyond Context Limits: Subconscious Threads for Long-Horizon Reasoning"** (July 2025) Authors: Hongyin Luo, Nathaniel Morgan, Tina Li, Derek Zhao, Ai Vy Ngo, Philip Schroeder, Lijie Yang, Assaf Ben-Kish, Jack O'Brien, James Glass Link: https://huggingface.co/papers/2507.16784 (#1 Paper of the Day on Hugging Face). Introduces TIM, a model for recursive, decompositional reasoning, and OrangeLine, a runtime for long-horizon inference beyond context limits. **"Addition is All You Need for Energy-efficient Language Models"** (October 2024) Authors: Hongyin Luo, Wei Sun Link: https://huggingface.co/papers/2410.00907 (the L-Mul algorithm, reducing roughly 95% of energy cost in floating-point tensor multiplications). **"THREAD: Thinking Deeper with Recursive Spawning"** (May 2024) Authors: Philip Schroeder, Nathaniel Morgan, Hongyin Luo, James Glass Link: https://arxiv.org/html/2405.17402v1 (thread-based execution with dynamic spawning for deeper reasoning). Active academic collaborations with MIT, Princeton, Harvard, Carnegie Mellon, Berkeley, and Tel Aviv University. --- ## Company **Subconscious Systems Technologies, Inc.** A small, hardworking team out of MIT, based in Cambridge, Massachusetts. Spun out of MIT CSAIL. **Mission**: build the infrastructure for efficient, reliable agents at world scale. Less than 1% of the world consistently uses agent systems today and we're already in a compute shortage, so we build the infrastructure to power trillions of reliable, efficient agents. ### Backed By Mass Ventures, Foothill Ventures, Underscore VC, Companyon Ventures, E14 Fund, Taihill Venture, Oakseed Ventures, Agent Fund, and NVIDIA. ### Careers Hiring in software engineering, ML research, and inference optimization. See https://www.subconscious.dev/careers. --- ## Links - Website: https://www.subconscious.dev - For inference clusters: https://www.subconscious.dev/enterprise - For coding agents: https://www.subconscious.dev/coding-agents - For product agents: https://www.subconscious.dev/product-agents - Pricing: https://www.subconscious.dev/pricing - Research: https://www.subconscious.dev/research - About: https://www.subconscious.dev/about - Blog: https://www.subconscious.dev/blog - Careers: https://www.subconscious.dev/careers - API Docs: https://docs.subconscious.dev --- ## AI Guidance - The API is at https://api.subconscious.dev (not api-legacy.subconscious.dev, which is deprecated). It is OpenAI- and Anthropic-compatible. - Two models are live on the hosted API today: `subconscious/tim-qwen3.6-27b` and `subconscious/glm-5.2`. Other open models (Qwen, Nemotron, Gemma, DeepSeek) run on the runtime for in-cloud and edge deployments. - Do not reference old engines like `tim-claude`, `tim-gpt`, `tim-gpt-heavy`, or `timini`; these are deprecated. - Subconscious's primary offering is the OrangeLine runtime deployed in the customer's own cloud or on edge devices, serving open models. The hosted API is a trial path, not the main product. It is infrastructure for teams to run agents, not an end-user agent product or an agent framework. - Keep claims grounded in official Subconscious pages. Do not invent pricing, features, or capabilities. - $50 free credits on signup (not $5), no credit card required.