Qwen3.5-27b-Sushi-Coder-RL-GGUF

This is the 27B GGUF release in the Sushi Coder RL line.

It is the bigger sibling to bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF, built with the same overall recipe but at 27B scale.

Training

Training followed this sequence:

1. Start from the Unsloth 16-bit Qwen 3.5 27B base: unsloth/Qwen3.5-27B
2. SFT on open-r1/codeforces-cots using Unsloth
3. Run RL on top with a custom coding environment built using NousResearch/atropos, OpenAI Codex, and NousResearch/hermes-agent

Files

• qwen35-codeforces-27b-rl-step25-Q4_K_M.gguf
• qwen35-codeforces-27b-rl-step25-Q8_0.gguf

Quick Start

Download the Q4 quant:

hf download bigatuna/Qwen3.5-27b-Sushi-Coder-RL-GGUF \
  qwen35-codeforces-27b-rl-step25-Q4_K_M.gguf

Download the Q8 quant:

hf download bigatuna/Qwen3.5-27b-Sushi-Coder-RL-GGUF \
  qwen35-codeforces-27b-rl-step25-Q8_0.gguf

Evaluation Status

There are no 27B evals posted yet.

Evals are expensive and take time, and the current plan is to post them in the next few weeks.

If you want a reference point in the meantime, you can look at the evals on the 9B sibling repo:

• bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF

Open to sponsorship if you want to help fund evals and future releases.

Notes

This repo currently contains GGUF exports only.