gpt-oss-20b

OpenAI released gpt-oss-120b and gpt-oss-20b on August 5, 2025 as open-weight language models under the Apache 2.0 license, trained using reinforcement learning and techniques from OpenAI's frontier systems including o3. The gpt-oss-20b variant delivers results comparable to OpenAI o3-mini on common benchmarks and can run on edge devices requiring only 16 GB of memory, making it well suited for local inference and on-device use cases. Both models support configurable reasoning effort levels, full chain-of-thought access, the Responses API, Structured Outputs, function calling, web search, and Python code execution for agentic workflows. OpenAI also reported strong tool use, few-shot function calling, CoT reasoning on the Tau-Bench agentic suite, and HealthBench results that outperformed proprietary models like o1 and GPT-4o.

A cross-vendor open-weights LLM release history and timeline published on hidekazu-konishi.com covers OpenAI's gpt-oss family alongside Llama, Mistral, Qwen, DeepSeek, and milestone references to Gemma and Phi. The article explicitly names gpt-oss-20b as part of the gpt-oss family and links release rows to official fir First published July 26, 2026 and last updated August 24, 2026, the timeline provides structured reference context for gpt-oss-20b's place in the broader open-weights landscape, useful for developers comparing model generations and licensing terms. The entry does not address OVHcloud-specific hosting or serving changes

An independent deployment analysis evaluated gpt-oss-20B on a single H100 GPU, measuring throughput, latency, VRAM, and energy against dense baselines Qwen3-32B and Yi-34B. The MoE model has approximately 20.9B total parameters with only 3.61B active per token (17.3% active fraction), delivering about 31.8% higher decode throughput and 25.8% lower energy per 1,000 generated tokens than Qwen3-32B at a 2048-token context. Peak VRAM was reduced by approximately 31.7% compared to Qwen3-32B, though time-to-first-token was higher due to MoE routing overhead. Normalized by active parameters, gpt-oss-20B shows markedly stronger per-active-parameter efficiency, underscoring MoE deployment advantages. The study did not evaluate accuracy, focusing solely on production deployment metrics.

OpenAI's gpt-oss-20B, released August 2025 as part of OpenAI's first open-weight LLM release since GPT-2, was evaluated against six open-source models across ten benchmarks covering reasoning, code generation, and multilingual tasks. The 20B variant consistently outperformed the larger gpt-oss-120B on several benchmarks including HumanEval and MMLU, while requiring substantially less memory and energy per response. Both gpt-oss variants demonstrated mid-tier overall performance within the open-source landscape, with relative strength in code generation and notable weaknesses in multilingual tasks. Results, validated using McNemar's test and effect size analysis, suggest scaling in sparse MoE architectures may not yield proportional performance gains. These findings inform more efficient model selection for open-source deployments.

The NVIDIA NIM modelcard for gpt-oss-20b describes a Mixture-of-Experts architecture with 21B total parameters and 3.6B active parameters, structurally identical to the 117B variant but with different hyperparameters. It uses SwiGLU activations and learned attention sinks, operates strictly on text input and output, and targets lower-latency local or specialized use cases. Highlights include the Apache 2.0 license for unrestricted commercial use, configurable low/medium/high reasoning effort, full chain-of-thought access for debugging, full parameter fine-tuning support, and native agentic capabilities including function calling, web browsing, Python code execution, and structured outputs.

The Baidu encyclopedia page on GPT-OSS-20B records that OpenAI released this open-weight model on August 5, 2025, with 21 billion total parameters and 3.6 billion active parameters per token, designed for low-latency local use on edge devices with 16GB of memory. It documents the model's Mixture-of-Experts Transformer The same page notes that GPT-OSS-20B's performance is reportedly comparable to OpenAI's o3-mini, with strong results on AIME and HealthBench, while some evaluations indicate its overall SWE-bench Verified performance is surpassed by Zhipu GLM-4.7-Flash. It also reports a 2026 study by institutions including The Univers

arXiv preprint 2508.16700 presents a single-GPU deployment evaluation of GPT-OSS-20B on NVIDIA H100 hardware in bf16 precision, explicitly naming gpt-oss-20b and characterizing it as a Mixture-of-Experts model with 20.9 billion total parameters and approximately 3.61 billion active parameters (about 17.3%). The study m At a 2048-token context with 64-token decode, the paper reports gpt-oss-20B delivers about 31.8% higher decode throughput, 25.8% lower energy per 1000 generated tokens, and 31.7% less peak VRAM than Qwen3-32B, while noting higher TTFT due to MoE routing overhead. It introduces an Active Parameter Efficiency (APE) metri