samuelcardillo/Carnice-Qwen3.6-MoE-35B-A3B

TEXT GENERATIONConcurrency Cost:3Model Size:35.1BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 19, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Carnice-Qwen3.6-MoE-35B-A3B by samuelcardillo is a 35.1 billion parameter Mixture of Experts (MoE) language model, with 3 billion active parameters per token, fine-tuned for agentic workflows and the Hermes Agent runtime. It is based on the Qwen3.6 architecture, offering a native context length of 262,144 tokens and native multimodal support. This model is specifically optimized for executing terminal commands, file operations, and multi-step tool calls within an agentic environment, making it ideal for complex automated tasks.

Loading preview...

Carnice Qwen3.6 MoE 35B-A3B: Hermes-Focused Agentic Model

Carnice-Qwen3.6-MoE-35B-A3B is a QLoRA fine-tune of the Qwen3.6-35B-A3B Mixture of Experts (MoE) model, developed by samuelcardillo. With approximately 35 billion total parameters and 3 billion active parameters per token, this model is specifically optimized for agentic workflows and the Hermes Agent runtime. It builds upon the Qwen3.6 base, which provides improved agentic coding capabilities, an extended native context length of 262,144 tokens, and native multimodal support.

Key Capabilities

  • Agentic Workflow Optimization: Trained on actual Hermes Agent execution traces, enabling native understanding of agentic conversation patterns.
  • Tool Use and Execution: Proficient in executing terminal commands, performing file operations, chaining multi-step tool calls, and utilizing browser-assisted workflows.
  • Decision Making: Capable of making decisions based on environmental feedback within agentic environments.
  • Extended Context: Features a native context length of 262,144 tokens, supporting complex, long-running agentic tasks.
  • Multimodal Support: Inherits native multimodal capabilities from the Qwen3.6 base model.

What Makes This Different

Unlike models trained on generic reasoning, Carnice-Qwen3.6-MoE-35B-A3B was fine-tuned using real Hermes Agent execution traces. This unique training approach teaches the model the precise conversational patterns and operational sequences expected by the Hermes Agent, including processing command outputs, editing files, and making informed decisions based on environmental interactions. This specialized training ensures highly effective and native agentic behavior.

Good For

  • Developing and deploying AI agents that interact with operating systems and tools.
  • Automating complex, multi-step tasks requiring environmental feedback and decision-making.
  • Applications requiring deep integration with the Hermes Agent runtime.
  • Scenarios demanding long context understanding for agentic reasoning and execution.

Finetunes
2 models