Haster1137/Qwen3.6-35B

Model Card: Qwen3.6-35B-A3B (Optimized for Agent Workflows) Model Details Developed by: Qwen (Alibaba Cloud) / Community Fine-tuned & Hosted

Architecture: Sparse Mixture-of-Experts (MoE) with Hybrid Attention

Total Parameters: 35 Billion

Active Parameters: 3 Billion per token (ultra-fast throughput)

Context Length: Up to 262K tokens

Primary Optimization: Designed for high-speed local inference, low-latency execution, and autonomous agent loops without corporate bloat or unnecessary safety friction.

Core Capabilities Autonomous Agent Execution: Built specifically for multi-step agent loops, file editing, debugging, and command execution.

Function & Tool Calling: Native, robust support for tool-use frameworks, custom API interactions, and structured outputs.

High-Throughput Performance: Thanks to its efficient MoE architecture (3B active parameters out of 35B total), this model delivers blazing-fast token generation speeds, making it ideal for continuous, heavy-duty coding and automation tasks.

Multimodal Support: Capable of processing text, code, and visual inputs across extended context windows.

Recommended Deployment & Usage Optimized to run on high-performance local hardware setups (such as NVIDIA Blackwell and Hopper architectures via vLLM) to achieve maximum token-per-second generation rates.

To serve with vLLM: vLLM serve Haster1137/Qwen3.6-35B --port 8000 --max-model-len 262144

This repository contains the weights for Qwen3.6-35B, a high-efficiency Mixture-of-Experts (MoE) model engineered for advanced reasoning, coding, long-context parsing, and agentic workflows.

Model Overview

• Architecture: Hybrid Gated DeltaNet + Gated Attention Mixture-of-Experts (MoE)
• Total Parameters: ~35 Billion
• Active Parameters per Token: ~3 Billion
• Context Window: Up to 262,144 tokens (262K native)
• Primary Strengths: Agentic coding, complex tool use, multi-step logical reasoning, and high-throughput generation

Key Features

1. Mixture-of-Experts Efficiency

By utilizing 256 fine-grained routing experts where only 8 fire per token, the model delivers the broad domain knowledge and reasoning depth of a 35B parameter backbone while operating with the speed and computational footprint of a 3B model.

2. Extreme Throughput

Thanks to sparse expert routing, this model achieves ultra-fast inference speeds (200+ tokens/second on high-memory enterprise hardware such as NVIDIA L40S, A100, or RTX PRO 6000 series), making it ideal for real-time applications and iterative agent loops.

3. Long-Context Capabilities

Designed with an extended 262K token context window, this model excels at large codebase analysis, long-form document parsing, and processing multi-turn conversational histories without losing track of instructions.

Recommended Usage

This model is compatible with standard OpenAI-compatible endpoints, TGI (Text Generation Inference), llama.cpp, vLLM, and Hugging Face Dedicated Inference Endpoints.

Recommended Generation Parameters

{
  "temperature": 0.7,
  "top_p": 0.8,
  "top_k": 20,
  "presence_penalty": 1.0
}