Subconscious is an AI lab solving the context window.
We research inference and post-training for long-horizon agents. Better caching makes better agents.
Tsinghua University
Inference
OrangeLine is our inference engine for agents, from how a request is routed to how its context is cached and compressed.
Global load balancer
Routing
Routes requests across our geographic clusters.
Cache-aware routing
Routing
Optimized for prefix and suffix caching, so each step of a trace lands where its context is already cached.
OrangeLine
Inference engine
Our inference runtime. It identifies stale spans, compresses them in the KV cache, reuses prefixes and suffixes, and drafts tokens with custom speculative decoders, so agents go further and faster on the same hardware.
L3 cache
Caching
A novel L3 cache built for long-horizon traces, so runs keep hitting cache deep into a trace.
Span identification
Scores each span of context as the run goes and finds the tool calls and results that stopped mattering.
Context compression
Compresses those spans in place on the GPU, freeing memory mid-run without a lossy summary.
Suffix caching
Keeps tokens on both sides of a compressed span cached, so later steps hit the cache instead of re-encoding.
Adversarial speculative decoding
Our own draft models propose tokens ahead of the main model, so generation stays fast deep into a run.
Subconscious Inference System
For long-horizon agents, 98% of tokens are cached tokens, so we've put most of our work into a better cache. That includes KV cache pruning with preserved suffix states: when context is pruned, what it contributed survives in the suffix instead of being re-encoded.
ASpec: Adversarial Speculative Decoding
Speculative decoding is usually treated as a pure speed win, but accepted draft tokens shift a model's output over long runs. ASpec trains the draft model to be rejected at critical decision points, so the target model samples those tokens itself. The target model stays unchanged, and long-horizon agents get more accurate from training 0.06% to 0.28% of hosted parameters.
Longer runs and more concurrency on the same GPUs.
The runtime manages memory and context during long agent runs.
Handle millions of tokens with context management at runtime.
With the highly efficient Subconscious Cache, cut cost 50% and up to 80%.
Run 2.3x as many workloads on the same compute footprint.
2x faster task completion down deep reasoning chains.
SGLang
Limited to the model context window with compaction. Long tasks run with unnecessary context or compact necessary information.
Subconscious
Processes millions of tokens per run. The system compresses its context at runtime, so accuracy stays high at any scale.
Post-training
RedLine is our post-training work, built on OrangeLine. We run SFT, OPD and RL on open-weight models, and post-train models for customers on their own data and tooling.
Compression at training time
Long reasoning trajectories train on significantly less compute.
Models that understand compression
They are optimized for compressed context and suffix reuse, so they get more out of OrangeLine.
Stronger small and mid-sized models
Higher accuracy on your own data, tooling and tasks.
Collaborate with us
We publish our findings on alphaXiv and work closely with academic labs. Email hongyin@subconscious.dev.