Summary: Guide Labs released the 8-billion-parameter Steerling-8B on February 23, 2026, with weights, code, and a package available as an architecture designed for interpretability. Its concept layer decomposes embeddings into known and discovered concepts, carries most token-level contribution, and supports tracing and inference-time steering without retraining. The company reports competitive 7B-scale results and about 90% of existing models’ capability, but those figures are not independently benchmarked and no license was named.
On February 23, 2026, San Francisco startup Guide Labs brought out Steerling-8B: an 8-billion-parameter base model designed from the start to be interpretable. The release covers three channels — weights on Hugging Face, code on GitHub, and a package on PyPI. TechCrunch’s label for it was “a new kind of interpretable LLM,” and the point of that phrase is that interpretability here is written into the architecture rather than recovered afterward by dissecting a trained network.
What changes for developers is practical: model behavior becomes an artifact you can inspect item by item and modify on the spot, instead of the subject of neuroscience studied through a sealed box.
Interpretability Inside the Architecture, Not Applied Afterward
Standard practice dissects a network after training has finished. Guide Labs CEO Julius Adebayo is blunt about the contrast: “The kind of interpretability people do is… neuroscience on a model, and we flip that.” During his MIT PhD he co-authored a widely cited 2018 paper demonstrating that the interpretability methods then available for deep learning were unreliable — the question he has kept coming back to ever since.
Mechanically, Steerling-8B places a concept layer inside the model itself. An embedding decomposes into approximately 33,000 supervised “known” concepts, roughly 100,000 “discovered” concepts the model arrived at on its own, and a residual component. Concepts reach the logits along a linear path, so each prediction splits exactly into the contribution of every concept. Steering follows from that same structure: at inference time you can add, drop, or combine concepts to change behavior, with no retraining at all.
Any Token Can Be Traced to Its Source
Whatever span of tokens the model generates, three threads run backward from it. One identifies which tokens in the input prompt did the influencing. Another shows which human-readable concepts were activated — tone concepts such as “analytical” or “clinical,” plus content concepts. The third pins down the training data sources in play: ArXiv, Wikipedia, FLAN, and others.
The measurements say this is structure, not decoration. Over 84% of token-level contribution passes through the concept module. Delete the residual pathway entirely and LM Harness results dip only slightly. Known-concept detection reaches 96.2% AUC on held-out validation data. Most of the model’s decision path travels through this layer.
What Interpretability Costs in Capability: Less Than Expected
The trade interpretable models have historically accepted is accuracy given up in exchange for transparency. Training for Steerling-8B consumed 1.35 trillion tokens, and the backbone is a causal discrete diffusion model, which allows steering across multi-token spans instead of being confined to next-token predictions. By Guide Labs’ reported numbers it beats LLaMA2-7B and Deepseek-7B on average while spending fewer FLOPs, and it remains competitive against models trained with 2–10x the compute and 2–7x the data.
Adebayo told TechCrunch that training interpretable models “is no longer a sort of science; it’s now an engineering problem,” and that nothing rules out such models matching frontier-scale performance. The company’s own claim is that Steerling-8B reaches roughly 90% of existing models’ capability with less training data.
One caveat deserves to stay in view: the performance figures come from the company. TechCrunch’s coverage includes no independent benchmark scores, and setting an 8B-scale model against LLaMA2-7B — a base model from 2023 — is a modest bar. So the honest claim is narrower than “interpretable models caught up”; it is “the transparency tax is no longer catastrophic at small scale.” Measured against where this field stood two years ago, that still counts as a real change.
Guide Labs and the Road Ahead
Guide Labs went through Y Combinator and closed a $9 million seed round in November 2024, led by Initialized Capital. Two founders run the company: CEO Julius Adebayo and Chief Science Officer Aya Abdelsalam Ismail. What’s queued up next covers a larger model, API and agentic access, and Clarity, an interpretable-AI platform running on an instruction-tuned version of Steerling-8B. The company’s blog also previews a run of deep dives — concept steering, concept discovery, alignment without fine-tuning (safety fine-tuning replaced by concept-level interventions), and memorization and training-data valuation.
Concrete uses are not hard to name: stopping copyrighted material from surfacing, keeping outputs away from violence or drug abuse, regulated industries in which a loan model ought to weigh financial records rather than race, and scientific work such as protein folding. The team also keeps track of “discovered” concepts the model found on its own — quantum computing among them — categories that arose during training with nobody supervising them.
What’s in Developers’ Hands Today
In hand today: weights on Hugging Face (guidelabs/steerling-8b), code on GitHub (guidelabs/steerling), and a PyPI package that installs directly. Teams working on output controls, model debugging, or compliance explanations get a complete reference implementation they can take apart. Two cautions apply: the training data itself did not ship with this announcement, and the post does not name a license — confirm both before commercial use.