SIQ-1-35B beats Qwen-350B with PPO-trained 35B fine-tune

SIQ-1-35B is a 35-billion-parameter fine-tune of Qwen-35B A3, trained with proximal policy optimization (PPO) and released this week on HuggingFace. The model outperforms GLM-5.2 and Qwen-350B on Karpathy's AutoResearch parameter-golf benchmark—a test that measures reasoning capability while minimizing parameter count. The developer reports this is the first PPO training run they've observed that delivered consistent improvements, attributed to a verifiable reward model rather than human preference signals.

On an internal "bullshit benchmark" used to measure output quality under adversarial prompting, SIQ-1-35B scores higher than NEX and GPT-5.5. The model also generates research ideas comparable in quality to Anthropic's Claude Opus, a closed frontier model typically reserved for complex reasoning tasks.

Weights are available in both full precision and GGUF quantized formats on HuggingFace. A free demo runs on HuggingFace Spaces with ZeroGPU backend, allowing users to test the model without API costs or local setup.

What stands out

1. 01PPO training that actually works. The developer emphasizes this is the first PPO run they've seen produce consistent improvements with a verifiable reward model rather than human preference signals.
2. 02Beats models 10× its size. SIQ-1-35B outperforms Qwen-350B (a 350-billion-parameter model) and GLM-5.2 on parameter-golf tasks, where the goal is to maximize capability while minimizing parameter count.
3. 03Opus-class idea generation. The model reportedly generates research ideas similar in quality to Anthropic's Claude Opus, a closed frontier model typically used for complex reasoning tasks.
4. 04Open weights and GGUF quantizations. Full weights and GGUF formats are available on HuggingFace, making the model accessible for local inference on consumer hardware.