auttasak88/AgentWorld-35B-A3B-Heretic

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

auttasak88/AgentWorld-35B-A3B-Heretic is a 35.1 billion parameter Mixture-of-Experts (MoE) language world model based on Qwen/Qwen-AgentWorld-35B-A3B, with approximately 3 billion active parameters per token. Developed by auttasak88 using the Heretic ablation technique, this model significantly reduces refusal and over-refusal behavior while preserving core capabilities like coding (HumanEval+ 85.4), instruction-following, reasoning, and multilingual support, including Thai. It is designed for research, agent-environment simulation, tool-use experimentation, and autonomous-agent system development, offering a less restrictive interaction experience.

Loading preview...

AgentWorld-35B-A3B-Heretic: A Refusal-Reduced Agent World Model

This model is a modified version of Qwen/Qwen-AgentWorld-35B-A3B, created by auttasak88 using the Heretic ablation method. Its primary goal is to substantially reduce refusal and over-refusal behavior while preserving the original model's core capabilities. Unlike typical tool-use agents, Qwen-AgentWorld models agent-environment interactions, simulating how environments respond to agent actions across domains like MCP, search, terminal, and software engineering.

Key Capabilities and Differentiators

  • Reduced Refusal: Achieves 83.2% HarmBench compliance and 0.0% XSTest over-refusal, a significant improvement over the base model.
  • Capability Preservation: Crucially, it maintains high performance in:
    • Coding: HumanEval+ (85.4) and MBPP+ (77.5) scores are preserved or slightly improved compared to the base model.
    • Instruction-Following & General Knowledge: IFEval and MMLU scores remain strong.
    • Multilingual Support: Thai language capabilities are well-preserved (e.g., Global-PIQA 86.0, MGSM-Thai 12.0), outperforming other ablated versions.
  • Architecture: Utilizes a qwen3_5_moe Mixture-of-Experts architecture with 35.1 billion total parameters, activating ~3 billion parameters per token for efficient inference.
  • Limitations: Some degradation is noted in GSM8K (math) and TruthfulQA, though less severe than in other ablated models.
  • Agent-Environment Simulation: While general capabilities are preserved, specific agent-environment simulation performance on domains like MCP, SWE, and Web (AgentWorldBench) is planned for future evaluation.

Deployment and Performance

  • Flexible Deployment: Supports llama.cpp (GGUF quantizations like Q4_K_M), vLLM (with --language-model-only and FP8 quantization for 2-GPU deployment), and Transformers.
  • Efficient Inference: Achieves ~186 tok/s single-stream on 1x RTX 3090 (llama.cpp Q4_0) and ~1,681 tok/s aggregate on 4x RTX 3090 (vLLM BF16, batch 64), benefiting from the MoE architecture's sparse activation.

Use Cases

This model is ideal for research, local inference, agent-environment simulation, tool-use experimentation, and autonomous-agent system development where a less restrictive and more compliant language model is desired without significant loss of core capabilities.

Finetunes
2 models