AstheralPlus is SEE7's instruction and reasoning model based on GPT-OSS-20B: 20.9 billion total parameters, with approximately 3.6 billion active per token.
The public checkpoint uses the original 128-step QLoRA run with 92,454,912 trainable adapter parameters, merged into native MXFP4 weights. Training sources are SmolTalk and OpenThoughts-114k.
Use the native Harmony chat template with low, medium or high reasoning effort. The Space provides Instruct and Thinking modes, project instructions, reference text files, saved chats and JSON export.
Training data · Training and evaluation · Native training experiments
Apache 2.0. Base model: openai/gpt-oss-20b.
model-00001-of-00002.safetensors 4.8 GBmodel-00000-of-00002.safetensors 4.8 GBmodel-00002-of-00002.safetensors 4.2 GBadapter/adapter_model.safetensors 370 MBopenai/gpt-oss-20b