Optitransfer/Qwen2.5-7B-Instruct-converge-collective-v1

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 18, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Optitransfer/Qwen2.5-7B-Instruct-converge-collective-v1 is a 7.6 billion parameter, training-free, weight-space merged model built upon Qwen2.5-7B-Instruct, designed to achieve top placements on public leaderboards. It features a 32,768 token context length and loads identically to its base model, using the same architecture, tokenizer, and chat template. This model demonstrates an improvement in GSM8K answer formatting, scoring 79.98% compared to 76.27% for the base model, indicating better adherence to expected output formats for mathematical problems.

Loading preview...

Optitransfer/Qwen2.5-7B-Instruct-converge-collective-v1 Overview

This model is the initial release from Optitransfer's converge program, a 7.6 billion parameter, training-free, weight-space merge built on Qwen2.5-7B-Instruct. It aims to integrate compatible fine-tunes from the open-source community into a single model to achieve top leaderboard placements. The model loads and prompts exactly like its base model, utilizing the same architecture, tokenizer, chat template, and configuration files.

Key Capabilities

  • Improved Answer Formatting: On the full GSM8K test set, v1 scores 79.98% (strict-match) compared to Qwen2.5-7B-Instruct's 76.27%, primarily due to a significant gain in producing answers in the expected #### N format.
  • Training-Free Merge: Constructed without any gradient training or additional training data, relying solely on weight-space merging techniques.
  • Identical Base Model Configuration: Retains the 32,768 token context length and other configurations of Qwen2.5-7B-Instruct.

Good For

  • Mathematical Reasoning (GSM8K): Demonstrates enhanced performance in mathematical problem-solving by consistently formatting answers correctly, which is crucial for automated evaluation systems.
  • Research and Lineage Study: Useful for researchers interested in the converge program's methodology and the evolution of merged models, particularly as it is superseded by v2.
  • Non-Commercial Research: Due to its lineage from facebook/opt-2.7b, this model is licensed for non-commercial research use only, making it suitable for academic and research-oriented projects.

Finetunes
1 models