GPT OSS 120B (Together AI)

SemiAnalysis InferenceX published a detailed architecture and evaluation breakdown of gpt-oss-120b, confirming it as an open-weight reasoning language model launched on August 5, 2025 alongside gpt-oss-20b under the Apache 2.0 license. The model is described as 117B total and 5.1B active parameters with a 131K token co The page highlights that over 90% of the model's weights are natively quantized to MXFP4 at roughly 4.25 bits per parameter, enabling the 120B model to fit on a single 80GB GPU such as an NVIDIA H100 or AMD MI300X. It also summarizes OpenAI's positioning claims, noting near-parity with o4-mini on core reasoning benchma

Databricks published a blog post demonstrating how its GEPA automated prompt optimization technique can lift gpt-oss-120b above leading proprietary frontier models on enterprise agent workloads. According to the post, gpt-oss-120b surpasses Claude Sonnet 4 and Claude Opus 4.1 by roughly 3% on the evaluation while being The same article shows that applying GEPA prompt optimization to the proprietary baselines raises Claude Opus 4.1 and Claude Sonnet 4 by an additional 6-7%, establishing new state-of-the-art results, and that prompt optimization matches or beats supervised fine-tuning while cutting serving costs by about 20%. Together