Independent research lab · Lagos
Most of the field treats efficiency as something you do to a model after it is finished. We train for it. What breaks when a model gets cheaper to run is rarely knowledge, it is behaviour, and it breaks quietly while the usual numbers keep looking fine. So we gate on behaviour and publish what the gate measured.
35.1B MoE in a 9.96 GB stock GGUF
Selected output
Models are one layer. We also build the runtime, memory, agent scaffolding, evaluation, and deployment path around them.
BTL-4 Compact keeps the model in a standard llama.cpp-compatible format while preserving 111 of 118 teacher-correct behaviors on the release gate. The paper records the range-selection result, expert-level ablations, and the boundary of the claim.
The frozen RL-0013 release combines agentic coding with structured tool use. It ships with complete evaluation evidence, a full-quality adapter, and a native compact edition.
OpenAI-compatible chat and responses APIs with provider routing, usage accounting, billing, caching, rate limits, and self-serve workspace keys.
A memory layer for agents that need to remember across sessions without turning every old fact into current truth.
Research with artifacts
Our research produces papers, datasets, environments, runtimes, model artifacts, and explicit failure reports—not just a thesis page.
CIB tests whether long-running agents preserve, update, retrieve, and act on the right evidence across sessions. The v0 release contains 250 deterministic tasks and an auditable evaluation pipeline.
R/02Echo-Skeleton Perception lets a text-only model operate a graphical interface through persistent structure, OCR, affordance probes, and typed change events instead of screenshots.
R/03BTL-4 Compact ships a 35.1B MoE as a 9.96 GB stock GGUF, while BRQ records the one-bit recovery lane and its non-promoted boundary.
R/04Controlled evaluations that separate observational pattern completion from interventional causal reasoning, with exact baselines and reproducible tasks.
Models, agents, infrastructure
Each project attacks a different failure mode: weak tool mechanics, forgotten context, unsafe autonomy, expensive inference, or perception that starts over every frame.
A local-first exoskeleton that gives small models explicit context, planning, tools, verification, critique, traces, and memory.
Open projectAn always-on lab operator with event ingestion, approvals, business loops, durable memory, founder briefs, and safe action boundaries.
Running inside BTLA privacy-first general agent with browser automation, files, shell, Python, scheduling, plugins, and messaging integrations.
Open project Open modelOur 7B code-review model produces structured security and correctness findings with file evidence and numeric confidence.
Open project Experimental runtimeA reusable Playwright and Tesseract implementation with stable screen nodes, event deltas, guarded actions, and external task verification.
Open project In developmentA local, ambient desktop agent built around selective attention, persistent context, permissioned action, and restraint.
Open projectTHE OPERATING PRINCIPLE
Build ambitious systems.
Measure them without mercy.
Ship what survives.