GPT-OSS-120B

What it does

GPT-OSS-120B is an open-source language model with 120 billion parameters, trained by distilling knowledge from DeepSeek V4 Flash with a focus on financial reasoning tasks. The model was distilled at an 8,000 token budget constraint. A 20 billion parameter version is available as open weights on Hugging Face. The project includes a playground for testing, evaluation datasets called LineageEval with 304 prompts, and code released on GitHub.

Who it is for

Developers and enterprises seeking open-source alternatives to proprietary models for finance-related tasks. The work appears aimed at American users concerned about dependencies on foreign models, particularly regarding potential behavioral transfer from censored Chinese models. The focus on financial reasoning suggests use cases in fintech, analysis, and related domains.

Pricing

The site does not show prices.

How it stands out

The model achieves 83.61% accuracy on FinanceReasoning benchmarks, exceeding Kimi K3 (81.93%) and Inkling (65.13%) at the same token budget. The project claims 62 times lower cost per query than Inkling and 160 times lower than Kimi K3. The core technical contribution examines whether censorship characteristics of the teacher model (DeepSeek V4 Flash) transfer to the distilled student model. The researchers found that despite training on outputs from a heavily censored Chinese model, the distilled version does not exhibit the same censorship behaviors on China-sensitive topics. This challenges assumptions that undesired behaviors necessarily transfer during knowledge distillation.

What a founder should check

First, verify the benchmark claims independently. The FinanceReasoning scores represent performance on specific tasks; test whether they generalize to actual production finance use cases your target users care about. Second, examine the cost comparison baseline. The claims about 62x and 160x cost advantages depend on how query costs are calculated and what comparable products actually charge in practice. Third, investigate the moat around open-source distilled models. If distillation from frontier Chinese models proves effective without behavioral transfer, competing teams can likely replicate this approach. The advantage may erode quickly as others adopt similar techniques. Check whether proprietary data, training methodology, or domain-specific refinements create defensible differentiation beyond the published benchmarks.