gpt-oss

TAS Score: S3/3 — D4/5 — A3 / T2 (D4 not D5: OpenAI attaches a short “gpt-oss usage policy” alongside the Apache-2.0 license — a minor overlay on otherwise OSI-clean weights. The license governing the weights is Apache-2.0, so transparency stays T2.)

Note: gpt-oss is OpenAI’s open-weight family (Apache-2.0), a separate object from OpenAI’s proprietary hosted models — see ChatGPT (A0/T0) for the same company’s closed offering.

Brief Description

OpenAI’s open-weight models — gpt-oss-120b and gpt-oss-20b, released August 2025 under Apache 2.0, the company’s first open-weight release since GPT-2. The 20b runs on a single high-end consumer GPU; the 120b needs a data-center card (~80 GB) or multi-GPU. Both are reasoning-oriented (configurable effort) and tool-use capable, self-hostable via Ollama, vLLM, and llama.cpp.

OpenAI attaches a brief “gpt-oss usage policy” alongside the license, but the license governing the weights themselves is OSI-approved Apache 2.0 — a clean T2. Training data, as usual, is not published. The pairing is the point: the same company that keeps GPT / ChatGPT fully closed (A0/T0) also ships gpt-oss fully open (A3/T2).

Architectural Role

Compute/inference layer: reasoning-capable open-weight models run on your own hardware. Local alternative to OpenAI’s hosted API for many tasks.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts (self-hosted)
  • Allows data export — Apache-2.0, standard HuggingFace format
  • OpenAI’s hosted GPT models are a separate paid API (not this card)

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Model weights stay on disk.
Exit ✅ Standard model format. Apache-2.0 — no restrictions on use or redistribution.
Recoverability ✅ Re-download from HuggingFace or restore from backup.
Visibility ✅ Apache-2.0 open weights. Brief usage-policy overlay. Training data not published.
External Dependencies ✅ Self-hosted: none. No first-party hosted dependency.

Configuration (Minimal) 

# Consumer GPU — 20b via Ollama
ollama run gpt-oss:20b

# Data-center GPU — 120b via vLLM
vllm serve openai/gpt-oss-120b --tensor-parallel-size 2

Alternatives

Alternative Autonomy Notes
Qwen3 A3 / T2 Apache open weights.
DeepSeek-R1 A3 / T2 Open-weight reasoning, MIT.
ChatGPT A0 / T0 Same company — closed hosted models.

Trajectory

Direction: opening

gpt-oss is OpenAI’s first open-weight release in years, under a clean Apache license — a notable opening from a company otherwise defined by closed APIs. Whether it becomes a sustained line or a one-off is the signal to watch; for now the released weights are permanently yours.

Signal assessment:

Signal Status Evidence
License ✅ Apache-2.0 on the weights (with a brief usage-policy overlay).
Feature gating ➖ Two sizes released; frontier hosted models remain separate and closed.
Self-hosting ✅ Standard formats. Ollama, vLLM, llama.cpp.
Governance ➖ OpenAI-controlled; open-weight commitment newly established, not yet proven durable.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

This site uses Just the Docs, a documentation theme for Jekyll.