Open AI GPT-OSS
GPT-OSS is OpenAI’s first open-weight model family since GPT-2. It comes in two text-only variants, gpt-oss-120b and gpt-oss-20b, released under an Apache-2.0 style license for local deployment, fine-tuning, and commercial use. Both target strong reasoning and “agentic” workflows (tool use, browsing, function calling) with adjustable reasoning effort.
What’s special about it (at a glance)
-
Open weights + permissive use: downloadable checkpoints on Hugging Face; supported across popular runtimes (vLLM, Ollama, Fireworks, etc.).
-
Runs on modest hardware: the 120B model uses Mixture-of-Experts (MoE) with ~5.1B “active” params per token and MXFP4 quantization so it fits on a single 80 GB GPU; the 20B model targets 16 GB devices.
-
Reasoning first: competitive math/coding and tool-use results and smooth “test-time scaling” as you dial up chain-of-thought length.
-
Safety work disclosed: extensive preparedness testing (cyber/biological) and jailbreak studies are summarized in the model card.
Comparison
| Model | SWE-bench Verified (%) | AIME 2025 pass@1 |
|---|---|---|
| GPT-OSS 20B (high) | 60.7 | 98.7% (with tools) |
| GPT-OSS 120B (high) | 62.4 | 97.9% (with tools) |
| GPT-4.1 (main) | 54.6 | — |
| OpenAI o3 (high) | 69.1 | 98.4% (with tools) |
| OpenAI o4-mini (high) | 68.1 | 99.5% (with tools) |
| GPT-5 (low) | — | — |
| GPT-5 (medium) | — | — |
| GPT-5 (high) | 74.9 | 94.6% (no tools) |
REF:
open weight
An open weight model is a type of AI model, particularly a LLM, where the final trained parameters, known as "weights," are made publicly available. These weights are the numerical values that have been learned by the model during its training on vast amounts of data and essentially represent the model's knowledge.
By releasing the weights, developers and researchers can download, examine, and use the pre-trained model to build their own applications, conduct further research, or fine-tune it for specific tasks.
No comments:
Post a Comment