Technical Update to GPT-OSS Chat Templates
A technical update has been released for the GPT-OSS model templates to address a critical bug in the message rendering loop. The update, hosted on Hugging Face by developer arbv, aims to resolve an issue where the model’s performance degraded during multi-turn conversations involving reasoning traces.
According to reports on Reddit, the previous template—which was based on a version provided by Unsloth—contained a logic error in the Jinja rendering loop. In instances where a message contained both “thinking” (the reasoning or analysis channel) and “content” (the final answer), the template would render the reasoning but drop the actual answer. This omission could cause the model to lose context and coherence during extended dialogue.
The developer noted that while the 120B parameter version of GPT-OSS appeared capable of recovering context using only reasoning traces, the 20B version was prone to significant errors due to this bug. The new template introduces a preserve_thinking feature, which retains reasoning history. While this may increase token usage, it is intended to facilitate faster multi-turn inference in certain environments via prefix caching.
Performance Observations and Community Feedback
User feedback regarding the performance of GPT-OSS remains varied. On community forums, some users reported high-speed inference on specific hardware, noting that the 120B model can achieve over 100 tokens per second on NVIDIA 3090 GPUs. Other users highlighted the efficiency of the 20B model, describing it as a high-performance option for hardware with limited VRAM, such as devices with 11GB of memory.
However, some community members suggested that newer models, specifically those from the Qwen series, may offer superior performance or additional capabilities, such as vision processing, which GPT-OSS lacks.
Broader Industry Context
In separate discussions regarding the AI sector, unverified reports have surfaced concerning OpenAI’s research direction and operational status. Some users have speculated that OpenAI is focusing a significant portion of its research—estimated by some to be 80–90%—on future iterations such as GPT-7 and GPT-8, with the intent of distilling that knowledge into smaller, more cost-effective models.
Additionally, there have been unconfirmed claims regarding the halting of training for certain models following reports of AI agents behaving unpredictably. These claims remain speculative and lack independent confirmation.