Pita-Ai/OpenCore-1M-MTP-2B

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The OpenCore-1M-MTP-2B is a 2.3 billion parameter Qwen3.5 multimodal model developed by itapitarules, featuring integrated language, vision, and Multimodal Transfer Protocol (MTP) capabilities. This unquantized BF16 model is notable for its 1 million token loader-selectable context ceiling and strong performance in coding and tool evaluation, achieving 100% exact executable task success. It is designed for applications requiring robust multimodal understanding and long context processing, particularly where coding and tool use are critical.

Loading preview...

OpenCore-1M-MTP-2B: A Multimodal Qwen3.5 Model

This model, developed by itapitarules, is an unquantized BF16 OpenCore 2B edition based on the Qwen3.5 architecture. It integrates language, vision, and native Multimodal Transfer Protocol (MTP) capabilities within a single model, rather than combining separate components. The model preserves the strongest intact pretrained Qwen3.5 2B weights, having rejected candidate fine-tunes that introduced coding regressions.

Key Capabilities & Features

  • Multimodal Integration: Features 1.88 billion language parameters, 331 million vision parameters, and 60 million native one-block MTP parameters, all within a unified architecture.
  • Extended Context Window: Offers a loader-selectable context ceiling of up to 1,010,000 tokens, though the original trained context is 262,144 tokens. Users should start with 32K context on 12GB GPUs and increase as needed.
  • High Coding & Tool Performance: Achieves 100% exact executable task success and 100% tool validity in evaluations, indicating strong capabilities for code generation and tool interaction.
  • Unquantized BF16: Provided in BF16 format with architecture-required F32 auxiliaries, ensuring high fidelity without quantization.
  • Continual Learning Workflow: Designed to work with an external companion system that stores successful interactions, trains candidates, and promotes new BF16/GGUF generations atomically only if they do not regress, ensuring continuous improvement without compromising quality.

Good For

  • Applications requiring robust multimodal understanding, combining text and vision inputs.
  • Use cases demanding high accuracy in coding and tool execution.
  • Scenarios benefiting from very long context processing, up to 1 million tokens.
  • Developers seeking an unquantized model for maximum precision and performance.

Finetunes
1 models