Glossary

OUR STACK

TIMRUNOUR RUNTIME

Our inference runtime, purpose-built for agentic workloads, and a drop-in alternative to vLLM and SGLang. It manages context aggressively during long runs: compressing the KV cache, reusing cached prefixes and suffixes, and freeing GPU memory mid-run. The result is 10x effective context, 2.3x concurrent workloads, and 3.5x token throughput compared to general-purpose engines on the same hardware. Run it on your own GPUs, or call our hosted API where open-source models run on TIMRUN.

Subconscious CacheFEATURE OF TIMRUN

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 TIMRUN

Lets TIM models prune their own reasoning history at inference time, inside the standard chat completions and messages formats. No server-side tool calls or special schemas. Subconscious Cache preserves memory across those prunes.

RedlineOUR POST-TRAINING SYSTEM

Our internal post-training system, built on top of TIMRUN. We use it to RL-train the TIM models so they exploit the runtime capabilities, like Auto Compaction and suffix reuse, that general-purpose training never sees. The runtime and the models are co-designed; Redline is the loop that ties them together.

TIMOUR MODEL FAMILY

Our family of post-trained open-weight models, built to run on TIMRUN and trained with Redline to manage their own context. The lineup spans device-sized variants up to flagship models. TIM-Qwen3.6 27B is live on our hosted API today.

Subconscious GLM-5.2LIVE ON OUR API

The open-source GLM-5.2 model, served on our TIMRUN runtime and built for agentic coding. It is our second supported model on the hosted API, alongside TIM-Qwen3.6 27B, and shows that any capable open model can run on TIMRUN, not just the post-trained TIM line.

THE ECOSYSTEM

Qwen

A family of open-weight models from Alibaba, and one of our favorite bases for post-training. TIM-Qwen models start from Qwen checkpoints before Redline tunes them for TIMRUN.

Nemotron

NVIDIA’s family of open models, and another base family we like to post-train. TIM-Nemotron variants are on the roadmap alongside the TIM-Qwen line.

GLM

A family of open-weight models from Zhipu AI (also known as Z.ai), strong on agentic and coding tasks. We serve GLM-5.2 on our hosted API today, running on TIMRUN — see Subconscious GLM-5.2 above — which shows that capable open models can run on our runtime, not just the post-trained TIM line.

Kimi

A family of open-weight models from Moonshot AI, built for long-context and agentic workloads. Kimi K2 is a large open-weight release in the line — the kind of capable open model that can be served on or post-trained for TIMRUN.

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 TIMRUN 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. TIMRUN 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 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 TIMRUN 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 TIMRUN 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. Our on-device offering is built on a llama.cpp-based runtime, so the same models can run on a laptop or workstation without a cloud round trip.