amayuelas/Qwen3.8-27B-MatRL-MT-RL

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 19, 2026Architecture:Transformer Featherless Exclusive Cold

The amayuelas/Qwen3.8-27B-MatRL-MT-RL is a 27.36 billion parameter Qwen3.8-based language model, fine-tuned with multi-turn reinforcement learning (RL) for novel inorganic crystal structure generation. It utilizes tool use against a machine-learned interatomic potential (MLIP) reward, making it specialized for materials science research. This model, with a 32768-token context length, is an intermediate checkpoint in a larger RL training pipeline, focusing on proposing and evaluating crystal structures.

Loading preview...

Model Overview

amayuelas/Qwen3.8-27B-MatRL-MT-RL is a specialized 27.36 billion parameter model built upon the Qwen3.8 architecture, designed for the generation of novel inorganic crystal structures. It leverages multi-turn reinforcement learning (RL) with tool use, guided by a machine-learned interatomic potential (MLIP) reward function.

Key Capabilities and Training

  • Crystal Structure Generation: The model proposes crystal structures in a multi-turn tool loop, submitting final candidates. It handles unconditional generation, target formula, target element set, and property-conditioned targets (e.g., band gap, bulk modulus).
  • Tool Use: Integrates various tools such as retrieve_materials, propose_structure, evaluate_structure, predict_band_gap, and submit to facilitate its task.
  • Reinforcement Learning: Trained with a GRPO-style framework, accumulating 170 RL steps on top of an initial 420 steps of cold-start supervised fine-tuning (SFT).
  • Reward System: Employs a weighted rubric considering submission, validity, novelty, stability (MLIP relaxation), property-target match, group diversity, efficiency, and format.
  • Context Length: Supports a max_model_len of 32768 tokens.

Important Considerations

  • Intermediate Checkpoint: This release is an intermediate snapshot (170 of a planned 400 cumulative RL steps) and is not a final, fully converged model.
  • Tool-Call Batching Drift: Users should be aware of a known issue where the policy progressively packs more tool calls into single assistant messages, potentially leading to truncated or malformed trailing tool calls if the 2048-token completion cap is hit.

Usage

This model is particularly suited for research and development in materials science, specifically for automated crystal structure discovery and optimization. It requires the multi-turn tool-calling format it was trained in, utilizing the bundled chat_template.jinja.