Here's an overview of the features we intend to work on in the near future. Feel free to comment on this issue below to suggest new directions for us to improve the library!

Fine-Tuning and Alignment

• Expand KerasHub as the premier model garden for fine-tuning by introducing high-level APIs for:
  • Model Distillation, for distilling capabilities from a large model to a smaller one;
  • Direct Preference Optimization (DPO) for preference tuning;
  • Reinforcement Learning: GRPO, PPO for mathematical reasoning, coding, etc.
• [Contributions Welcome] DoRA (Weight-Decomposed Low-Rank Adaptation)

High-Performance Inference and Usability

• Improve the generate() method for LLMs, focusing on memory reduction and long-context optimizations by exploring techniques like KVPress, Speculative Decoding, etc.
• Integrate vLLM to provide a backend for state-of-the-art inference throughput and performance.
• Introduce support for creating multi-turn conversational experiences (chat with your models).
• Safetensor export API.
• [Contributions Welcome] Introduce a high-level API for popular speech-to-text models like Whisper.

Models

Continue to rapidly expand the collection of pretrained models available on KerasHub.

• New Gemma models: Gemma 3n and Gemma T5.
• Vision-Language Models (VLMs): SmolVLM, Qwen-VL, SmolVLM2, and others.
• Leading Open Source Models: Phi-3/4, Qwen3 MoE, StableLM-3B, Deepseek, Magistral, ModernBERT, Kimi K2.
• Domain-Specific Architectures: Models like ESM for biology and LayoutLMv3 for document AI.

Integrations and Ecosystem

• Integrate Tunix, a new JAX-based library for LLM post-training - RL, Distillation, etc.