amayuelas/Qwen3.5-9B-MatRL-MT-RL

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The amayuelas/Qwen3.5-9B-MatRL-MT-RL is a 9 billion parameter Qwen3.5-based model developed by amayuelas, specifically fine-tuned for multi-turn reinforcement learning in crystal-structure inverse design. It proposes, evaluates, and refines candidate structures against machine-learning interatomic potentials. This model excels at tasks involving material science inverse design, optimizing for validity, novelty, thermodynamic stability, and property targets within a 32,768 token context window.

Loading preview...

Model Overview

amayuelas/Qwen3.5-9B-MatRL-MT-RL is a 9 billion parameter model built upon amayuelas/Qwen3.5-9B-MatRL-MT-SFT, specialized for multi-turn reinforcement learning (RL) in crystal-structure inverse design. This model functions as an agent that proposes and refines candidate structures, evaluating them against machine-learning interatomic potentials. Its primary objective is to optimize for validity, novelty, thermodynamic stability, and specific property targets like band-gap and bulk-modulus.

Key Capabilities & Training

  • Crystal Structure Inverse Design: Optimized for generating and refining crystal structures through an iterative RL process.
  • Multi-turn RL: Utilizes a GRPO-style multi-turn RL algorithm with 10 turns and 8 evaluations per episode.
  • Comprehensive Reward System: Rewards are based on a combination of factors including SUN-marginal (validity, novelty, relaxed stability using equflashv2 MLIP), group diversity, and specific property targets (band-gap, bulk-modulus).
  • Flexible Task Mix: Training includes 20% formula-conditioned, 10% unconditional, 10% band-gap-target, and 10% bulk-modulus-target prompts.
  • Extended Context Window: The model was trained with a sequence length of up to 32,768 tokens in its final stages.
  • Native Thinking Channel: Designed to preserve reasoning across tool calls, requiring enable_thinking=true for optimal performance.

Use Cases

This model is particularly well-suited for:

  • Materials Science Research: Automating the design and discovery of novel crystal structures with desired properties.
  • Computational Chemistry: Exploring the vast space of possible material compositions and configurations.
  • AI-driven Material Discovery: Accelerating the inverse design process for new materials with specific thermodynamic or electronic characteristics.

Finetunes
1 models