On February 23, 2026, San Francisco startup Guide Labs released Steerling-8B: an 8-billion-parameter base model designed from the ground up to be interpretable. The weights are on Hugging Face, the code is on GitHub, and there is a PyPI package. TechCrunch called it “a new kind of interpretable LLM” — not a trained network dissected after the fact, but a model with interpretability written into its architecture.
For developers, the significance is straightforward: model behavior becomes something you can inspect line by line and edit directly, rather than neuroscience performed on a black box.
Interpretability Built Into the Architecture, Not Bolted On
The conventional approach dissects a finished network. Guide Labs CEO Julius Adebayo puts it bluntly: “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 showing that existing deep-learning interpretability methods were unreliable — the problem he has been circling ever since.
Steerling-8B inserts a concept layer into the model itself. Embeddings decompose into roughly 33,000 supervised “known” concepts, about 100,000 “discovered” concepts the model learned on its own, and a residual component. Concepts feed into the logits through a linear path, so every prediction decomposes exactly into per-concept contributions — and at inference time you can add, remove, or compose concepts to steer behavior without any retraining.
Every Token Traces Back
Any group of generated tokens can be traced in three directions: which tokens in the input prompt influenced it, which human-readable concepts it activated (tone concepts like “analytical” or “clinical,” plus content concepts), and which training data sources it draws from — ArXiv, Wikipedia, FLAN, and others.
The numbers make the case that this is structural, not cosmetic. More than 84% of token-level contribution flows through the concept module; remove the residual pathway entirely and LM Harness results barely move; known-concept detection hits 96.2% AUC on held-out validation data. The concept layer carries most of the model’s decision path.
The Capability Cost Is Smaller Than Expected
The historical bargain for interpretable models has been transparency in exchange for accuracy. Steerling-8B was trained on 1.35 trillion tokens, on a causal discrete diffusion backbone that allows steering over multi-token spans rather than only next-token predictions. Guide Labs reports it outperforms LLaMA2-7B and Deepseek-7B on average while using fewer FLOPs, and stays competitive with models trained on 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 there is no reason such models cannot match frontier-scale performance. The company claims Steerling-8B reaches roughly 90% of existing models’ capability with less training data.
One caveat worth holding onto: these performance figures are company-reported. TechCrunch’s coverage cites no independent benchmark scores, and 8B-scale comparisons against LLaMA2-7B — a 2023-era base model — set a modest bar. The honest claim is narrower than “interpretable models caught up”: it is “the transparency tax is no longer catastrophic at small scale,” which is still a meaningful change from where this field stood two years ago.
The Company and Its Roadmap
Guide Labs is a Y Combinator alum that raised a $9 million seed round led by Initialized Capital in November 2024. Its two founders are CEO Julius Adebayo and Chief Science Officer Aya Abdelsalam Ismail. Next on the plan: a larger model, API and agentic access, and Clarity, an interpretable-AI platform powered by an instruction-tuned version of Steerling-8B. The company’s blog previews a series of deep dives — concept steering, concept discovery, alignment without fine-tuning (replacing safety fine-tuning with concept-level interventions), and memorization and training-data valuation.
The application space is concrete: blocking copyrighted material, controlling outputs related to violence or drug abuse, regulated industries where a loan model should weigh financial records rather than race, and scientific use cases like protein folding. The team also tracks “discovered” concepts the model found on its own — quantum computing, for instance — categories that emerged during training without supervision.
What Developers Get Today
The weights live on Hugging Face (guidelabs/steerling-8b), the code on GitHub (guidelabs/steerling), with a PyPI package for direct installation. For teams building output controls, debugging model behavior, or producing compliance explanations, this is a complete reference implementation you can take apart. Two caveats: the training data itself was not released with this announcement, and the license is not specified in the post — verify both before commercial use.
Sources
- Steerling-8B Base Model Release — Guide Labs
- Guide Labs debuts a new kind of interpretable LLM — TechCrunch
- guidelabs/steerling-8b — Hugging Face
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.