Research

Subconscious is an AI lab making open models far more capable.

We build TIMRUN, an inference runtime designed for agentic work, and the complementary post-trained TIM family of models.

Pillar 1: Inference Runtime

An inference system designed for agents

General-purpose inference systems waste compute on agent workloads. We built TIMRUN, a runtime that treats agent traces as the primary workload for better performance and reliability.

Featured · June 2026

Subconscious Cache

Our system for KV cache pruning with preserved suffix states. When a span of context is pruned, the information it contributed survives in the retained suffix states instead of being thrown away and re-encoded.

50%

Lower latency per task

2.3×

Supported concurrent workloads

3.5×

Faster token generation

Read the post

Auto Compaction

Our runtime enables models to prune their own message list at inference time. Compaction happens automatically inside the familiar chat completions and messages APIs, and the Subconscious Cache preserves memory across those prunes.

Read the post

TIMRUN in Production

This is more than research, it's live in production today. Customers utilize our inference runtime via API, on their own GPU clusters, and on local devices.

Read the post

STANDARD RUNTIME · PREFIX CACHE ONLY

System + tools
Pruned
Recent steps
Output

Everything after the pruned span is re-encoded from scratch.

SUBCONSCIOUS TIMRUN · PREFIX + SUFFIX CACHE

System + tools
Pruned
Recent steps
Output

The suffix is reused too. Some of the pruned token information is preserved in the suffix.

Cache hitRe-encodedPruned by the agent

Each strip is one agent context after a mid-conversation prune. A standard prefix cache only reuses tokens before the pruned span, so everything after it is re-encoded. TIMRUN also reuses the suffix, keeping most of the context cached.

Pillar 2: Model Post-training

Redline

TIMRUN alone is powerful. Redline is how open models learn to use it even better.

01

Co-designed with the runtime

Redline uses TIMRUN under the hood and trains the models to exploit context compression and suffix reuse.

02

Highly efficient post-training

With context compression during training, Redline requires up to 75% less compute to train on huge reasoning trajectories. That cost saving flows straight through to our customers.

03

SFT, OPD, and RL on agentic workloads

We use it to run SFT, OPD, and RL on open-weight models on long-horizon, tool-heavy tasks. Coding agents, browser control, workflow automation, research tasks, and more.

04

For our TIM models, or yours

Redline is how we create our TIM family of models, but also how we post-train models for customers on their unique data and tooling.

Born at MIT CSAIL

Publications

The work above stands on research we published in the MIT ecosystem.

Jul 2025

#1 Paper of the Day on Hugging Face

Beyond Context Limits: Subconscious Threads for Long-Horizon Reasoning

The paper that started it all. Introduces TIM, a language model designed for recursive, decompositional reasoning, and TIMRUN, a runtime that prunes irrelevant context in real time — enabling long-horizon inference beyond GPU and positional-embedding limits. Everything on this page grew out of this work.

Hongyin Luo, Nathaniel Morgan, Tina Li, Derek Zhao, Ai Vy Ngo, Philip Schroeder, Lijie Yang, Assaf Ben-Kish, Jack O'Brien, James Glass

Read paper

Jan 2025

Beyond Single-Shot: Multi-step Tool Retrieval via Query Planning

TOOLQP models tool retrieval as iterative query planning: instructions are decomposed into sub-tasks and queries are generated dynamically, trained with RL from verifiable rewards. State-of-the-art retrieval with strong zero-shot generalization and better downstream agentic execution.

Wei Fang, James Glass

Read paper

Oct 2024

Addition is All You Need for Energy-efficient Language Models

L-Mul approximates floating-point multiplication with integer addition — higher precision than 8-bit floating point at far lower cost, cutting up to 95% of the energy of element-wise tensor multiplications and 80% of dot products.

Hongyin Luo, Wei Sun

Read paper

May 2024

THREAD: Thinking Deeper with Recursive Spawning

Frames model generation as a thread of execution that can dynamically spawn child threads, offloading work and returning only the tokens the parent needs. The model adapts how much intermediate work it does, reaching state-of-the-art performance on diverse benchmarks.

Philip Schroeder, Nathaniel Morgan, Hongyin Luo, James Glass

Read paper

Apr 2023

Interpretable Unified Language Checking

UniLC shows LLMs are inherent multi-task language checkers: one interpretable method checks whether human or machine-generated language is factual and fair, outperforming fully supervised baselines on several tasks.

Tianhua Zhang, Hongyin Luo, Yung-Sung Chuang, Wei Fang, Luc Gaitskell, Thomas Hartvigsen, Xixin Wu, Danny Fox, Helen Meng, James Glass

Read paper

Open collaboration

We publish, and we partner.

Our research is informed by our connections to the academic community and our commitment to publishing our findings. Today we collaborate with labs at MIT, Princeton, Harvard, Carnegie Mellon, Tel Aviv, and Berkeley.

For research collaborations, academic partnerships, or questions about our work, reach out to hongyin [at] [this website domain].

© 2026 Subconscious Systems Technologies, Inc.

Subconscious