AI Infra Interviews logo

gpt-oss / Lab profile

OpenAI

The older open release remains a useful compact reasoning baseline.

First time comparing models? Learn the terms before the architecture →

The August 2025 gpt-oss pair remains relevant because it exposes a practical small-versus-large reasoning comparison within one release. These are downloadable text models; they should not be described as the weights behind current hosted GPT services.

Browse the 2 covered releases ↓Official lab website ↗

Architecture trajectory

What connects these models?

Both models use sparse experts and mixed MXFP4 expert storage. They combine local and full attention, and rely on the Harmony conversation format. Adjustable reasoning effort changes the output process and therefore latency, not just an abstract quality setting.

Local and global layers read different spans.

A token enters a layerthe layer type selects its spanLocal attention layerread a bounded token windowGlobal attention layerread across available contextWindowed KV statebudget follows the local windowGlobal KV statebudget grows with context

Window length and layer ratio are checkpoint-specific.

Full-context state remains for global layers.

Schematic of a sliding-window and full-attention stack, not a literal layer ordering. Local layers attend within a bounded token window; global layers attend across the available context. These are both attention layers, not a recurrent-state architecture.

Newest first / grouped sequence, not a time scale

The gpt-oss timeline

  1. gpt-oss-120b

    A single large-memory GPU is a useful deployment class

    It paired 117B total parameters with 5.1B active per token in the same August 2025 release as 20b. The model demonstrates how artifact design can change deployment class without changing parameter count.

    Release evidence ↗

  2. gpt-oss-20b

    A compact reasoning baseline that still earns its place

    The August 2025 release brought open-weight reasoning, Harmony formatting and mixed MXFP4 expert storage together. Those integration choices remain relevant even when you select a newer model.

    Release evidence ↗

Which other labs should you compare?

Compare 20b with 120b to isolate scale under the same release and formatting rules. GLM-4.7-Flash is a newer small MoE alternative; Qwen or Gemma is a better comparison when visual input is a requirement. Recency alone should not disqualify a model that meets the workload cheaply.

GLM

Z.ai

Track the attention changes separately from the post-training gains.

Qwen

Alibaba Qwen

One family spans tiny dense models, coding MoEs and conditional memory.

Interview preparation

What you should be able to explain

Explain why 117B parameters need not imply 234 GB of deployed weights. Then connect reasoning effort to output-token count, latency and task quality.

Study the underlying systems → · Get the illustrated report →

Sources and scope

Checked 13 September 2026. Assessments and comparison candidates are our interpretation of the cited designs, not measured cross-lab rankings. Each timeline entry links to its release evidence; model pages preserve pinned card and configuration sources.