Skip to content

Hibou-B / Hibou-L

Family: path-fm Modality: H&E and non-H&E whole-slide image tiles (224×224, patch size 14); cytology slides included in pretraining Released: 2024 (Nechaev, Pchelnikov, Ivanova — HistAI; arXiv June 2024) License: Apache-2.0 for both Hibou-B and Hibou-L — commercial use permitted (one of the only major pathology FMs with this property) Code/checkpoint: github.com/HistAI/hibou; HuggingFace histai/hibou-b, histai/hibou-L Also surfaced in: single-cell-spatial-omics

Hibou is a family of two self-supervised pathology foundation models from HistAI (Hibou-B, 86M parameters; Hibou-L, 307M parameters) that uses a register-augmented DINOv2 recipe to produce general-purpose tile embeddings for H&E and non-H&E histology. Unlike most other pathology FMs at this scale, both Hibou variants ship under Apache-2.0 with no commercial restriction, which has made them a common backbone for industry-adjacent and startup pipelines that cannot adopt UNI / Virchow-2 / Prov-GigaPath under their respective non-commercial terms.

Architecture & training

Hibou-B is a ViT-B/14 with ~86M parameters; Hibou-L is a ViT-L/14 with ~307M parameters. Both use a customized DINOv2 implementation that adds register tokens (Darcet et al. 2024) on top of the original Facebook Research code — register tokens stabilize self-supervised ViTs at scale by absorbing the "high-norm artifact" tokens that otherwise pollute attention maps. The implementation requires trust_remote_code=True to load from HuggingFace.

Pretraining corpus. The training set comprises 1,138,905 whole slides: 936,441 H&E-stained slides, 202,464 non-H&E (mostly IHC and special stains), plus 2,676 cytology slides, drawn from 306,400 unique cases. The corpus is multi-institutional and includes veterinary biopsies alongside human tissue — the authors frame this as one of the most stain- and species-diverse pathology pretraining sets to date.

Compute and recipe. Hibou-B was trained on 512M curated patches using 8× A100-80G GPUs with batch size 1024 for 500k iterations. Hibou-L was trained on 1.2B curated patches using 32× A100-40G GPUs, batch size 1024, for 1.175M iterations. Augmentations included random rotation, horizontal/vertical flips, RandStainNA stain normalization, and color jittering. The training objective is standard DINOv2 (DINO + iBOT + KoLeo) with register tokens.

Hibou is a tile encoder only — no slide-level head — and is intended to feed attention-MIL or linear probes for downstream tasks.

Headline benchmarks

Patch-level linear probing (Hibou paper, average top-1 across 6 public datasets: CRC-100K, PCAM, MHIST, MSI-CRC, MSI-STAD, TIL-DET).

Model Avg. top-1
Phikon 0.846
Virchow 0.871
RudolfV 0.871
Hibou-B 0.872
Hibou-L 0.887

Hibou-L is the top average performer in this comparison and beats Virchow by ~1.6 absolute points, despite Virchow having roughly 2× the parameters.

Slide-level AUC (Hibou paper, weakly-supervised aggregation). Hibou-L reaches 0.946 on TCGA-BRCA subtyping, 0.969 on TCGA-NSCLC, and 0.996 on TCGA-RCC. Notably, Hibou-B (86M) surpasses Prov-GigaPath (1.1B tile encoder) on 2 of the 3 slide-level benchmarks despite having ~13× fewer parameters — the authors cite this as evidence that pretraining-data diversity (non-H&E + cytology) contributes meaningfully alongside scale.

Independent benchmarks. Public ranking suites of pathology FMs (e.g. the 2025 ArteraAI clinical readiness review) typically place Hibou-L mid-tier on H&E-only tasks — below UNI2-h and Virchow-2 on average accuracy, but the only model in that tier with a permissive commercial license.

Why it matters for AACR / clinical pathology

  • Apache-2.0 is the differentiator. UNI / UNI2-h, CHIEF, and Virchow-2 all gate weights under CC-BY-NC or AGPL non-commercial terms; Prov-GigaPath weights are gated under Microsoft Research / Providence terms. Hibou is the only top-tier pathology FM with unrestricted commercial weights, so it is disproportionately represented in startup and pharma-adjacent pipelines that need a clean license trail.
  • Non-H&E coverage. The 200K+ IHC and special-stain slides in pretraining make Hibou competitive on multi-stain tasks where UNI / Virchow-1 (H&E-only) underperform. This matters at AACR for posters combining H&E with companion-diagnostic IHC (HER2, PD-L1, Ki-67).
  • Lightweight option. Hibou-B at 86M parameters is the most inference-cheap of the headline pathology FMs, useful for whole-cohort scale-out and on-prem deployment.

Known limitations

  • No slide-level encoder. Like Virchow and UNI, Hibou is a tile encoder; users must add their own MIL aggregator.
  • Smaller training corpus than Virchow-2 / Prov-GigaPath. 1.1M slides is below Virchow-2 (3.1M) and Prov-GigaPath (171K slides but 1.3B tiles); on pure H&E tasks the larger models still lead.
  • Closed pretraining data. The dataset is proprietary to HistAI; no slide-level provenance, demographic breakdown, or fairness audit is published.
  • Smaller benchmark suite in the paper. Hibou's headline comparisons cover 6 patch datasets and 3 TCGA slide tasks — less exhaustive than the 26+ tasks in Prov-GigaPath or the 30+ in UNI.

Sources

  • Nechaev D., Pchelnikov A., Ivanova E. (2024). Hibou: A Family of Foundational Vision Transformers for Pathology. arXiv:2406.05074.
  • HuggingFace: histai/hibou-b (Apache-2.0), histai/hibou-L (Apache-2.0).
  • GitHub: HistAI/hibou — reference code, training notes, and downstream examples.
  • Darcet T. et al. (2024). Vision Transformers Need Registers. ICLR 2024 — methodological basis for Hibou's register-augmented DINOv2.

Use in the AACR 2026 corpus

Mention scan pending — populated by the alias-matching pipeline used for UNI, CHIEF, and Prov-GigaPath.