Skip to main content
Back

Qwopus3.6 Beats Q8 Quality at Half the Size on Consumer Hardware

Qwopus3.6-35B-A3B-Coder gets APEX quantization and bundled MTP heads for faster, leaner local coding agents

  • New release: Qwopus3.6-35B-A3B-Coder-APEX-MTP-GGUF combines APEX quantization and a bundled MTP head for faster local inference.
  • APEX quantization assigns different precision per tensor role in MoE models, beating Q8_0 quality at roughly half the file size.
  • Self-speculative decoding is enabled with a single --draft-mtp flag -- no separate draft model required, files are only ~2.5% larger.
  • 62.4% on SWE-bench (300 cases, thinking-off mode, Q5_K_M quant) -- strong results for a local model running without chain-of-thought.
  • Five quantization tiers from 21.3 GB (I-Quality) to 12.2 GB (Mini) cover everything from high-accuracy to consumer GPU deployment.
  • Free and open-source under Apache 2.0; best for OpenHands, OpenCode, and Codex-style local coding agent workflows.

Qwopus3.6-35B-A3B-Coder-APEX-MTP-GGUF is a new community release that combines two independent innovations into a single file: a smarter quantization scheme purpose-built for Mixture-of-Experts models, and a bundled multi-token prediction head that enables self-speculative decoding without any extra setup. The result is a 35B-parameter coding agent that fits on consumer hardware and runs faster than standard quantized builds, with no quality trade-off.

A model built for doing, not thinking

Qwopus-3.6-35B-A3B-Coder is a thinking-off, token-efficient coding agent model designed for real agentic coding workflows where the model repeatedly reads files, chooses tools, edits code, runs tests, reacts to errors, and summarizes work. The core goal is to complete more of these steps with less token waste, lower latency, and more stable behavior when explicit long thinking is disabled.

Most coding-focused LLMs are optimized to reason longer and harder. This one takes the opposite bet: that most steps in a real agent loop (pick a tool, read a file, patch a function) don't need deep deliberation. They need speed and consistency. The model is optimized for repeated tool decisions, patching, test runs, and error-driven debugging without forcing every step into long thinking mode.

The architecture underneath

The coder model builds on the Qwopus3.6-35B-A3B line, itself based on Qwen3.6-35B-A3B. The underlying architecture is a hybrid sparse MoE model with 35B total parameters and approximately 3B active parameters per token, making it attractive for local high-frequency coding workloads. Only 3B parameters fire per forward pass, which keeps inference fast even on a single GPU.

The training workflow is accelerated and memory-optimized with Unsloth. The fine-tuning pipeline layered coding and tool-use adaptation on top of the base MoE, then specifically targeted thinking-off behavior: faster next-step decisions, lower token waste, and better workflow stability across long agent loops.