Keras ships ZeroModels: 100+ model families in pure Keras 3, runnable on any backend
fchollet · x · 2026-09-10
François Chollet announced ZeroModels: 100+ model families implemented in pure Keras 3, with weights converted from original checkpoints. The same code runs on JAX, PyTorch, and TensorFlow — no transformers or torch dependency at runtime.
Coverage includes object detection (DETR), segmentation (SegFormer, 13 variants), depth estimation (Depth Anything V2), promptable masks (SAM 3), open-vocabulary detection (OWLv2), and speech recognition (Whisper), with SAM 3, Gemma 4, and GLM-5 newly landed.
Usage is uniform: fromweights builds both the model and its processor from one Hub identifier, so moving between a detector, a depth estimator, and an LLM costs nothing. Install via pip install -U zeromodels. Combined with KerasHub, the Keras ecosystem now spans an extensive range of foundation models.
Related event: Keras Launches ZeroModels: 100+ Model Families Run Across Backends(2 posts)→
More from Infra
- Adaptive VRAM governor keeps 130M model training for 1M steps on an 8GB GPU without OOM — uBazzyZ- · 2026-09-10
- Epoch: Top AI firms' compute grows 4x/year; OpenAI up nearly 20x since 2023 — Jsevillamol · 2026-09-10
- Qwen3.8 Flash on 128GB Strix Halo: memory math and perplexity of Q4+Q8 n-gram hybrid quants — MarkoMarjamaa · 2026-09-10
- Agent auto-builds a custom Dockerfile to optimize new workspace start times — lucasmeijer · 2026-09-10
- OpenAI compute has grown ~20x since 2023; leading labs scale 4x/year — The Verge AI · 2026-09-10
- Cohere moves to NVIDIA Blackwell, cutting token costs and TTFT by 30–50% — cohere · 2026-09-10