GestaltLabs/Ornstein-3.5-9B-V2

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

GestaltLabs/Ornstein-3.5-9B-V2 is a 9.65 billion parameter language model developed by GestaltLabs, building upon Ornstein 3.5 9B V1.5 with DPO and GRPO verifiable-reward reinforcement learning. This model features a native vision tower and multi-token-prediction head, and is specifically optimized for graduate-level science (GPQA) and multi-step reasoning tasks. It excels in AI-research assistance and technical problem-solving, achieving ceiling performance on reasoning benchmarks.

Loading preview...

Ornstein 3.5 9B — V2 Overview

GestaltLabs/Ornstein-3.5-9B-V2 is a 9.65 billion parameter model developed by GestaltLabs, representing the second version of the Ornstein 3.5 9B series. It builds upon the V1.5 supervised fine-tune by incorporating advanced reinforcement learning techniques, specifically DPO (preference optimization) and GRPO (verifiable-reward reinforcement learning on math RLVR). This post-training sharpens the model's capabilities, particularly in complex reasoning and scientific understanding.

Key Capabilities & Features

  • Enhanced Reasoning: Achieves ceiling performance (1.00) on the Gestalt Benchmark Suite (GBS-200) for reasoning and GPQA (graduate-level science) tasks, significantly outperforming its base model (Qwen3.5-9B-Base).
  • Multimodal: Retains the native vision tower for image and video input via the bundled mmproj.
  • Speculative Decoding: Includes a multi-token-prediction (MTP) head for efficient speculative decoding.
  • Post-training: Utilizes DPO and GRPO verifiable-reward RL to refine its output based on reward signals.
  • Quantizations: Available in various formats including GGUF (with mmproj), bf16 GGUF, AWQ int4, and NVFP4.

Intended Use Cases

  • Reasoning-heavy tasks: Ideal for problems requiring deep logical inference.
  • AI-research assistance: Supports complex research queries and problem-solving.
  • Technical and scientific problem-solving: Excels in graduate-level science and multi-step reasoning.
  • General conversation: Capable of engaging in broad conversational contexts.

Finetunes
4 models