GPT OSS 20B (Deep Infra)

GPT-OSS-20B is an open-weight language model built on a Mixture-of-Experts architecture that activates only a fraction of its weights per token. According to an independent deployment-focused study, the model carries roughly 20.9B total parameters but engages only about 3.61B of them at inference, leaving the active parameter share at around 17.3 percent. This sparse-activation design is the core of its identity and informs both its efficiency profile and the routing-related overhead observed during prefill.

An independent single-GPU benchmark on an NVIDIA H100 in bf16 precision found that GPT-OSS-20B delivers higher decode throughput and more tokens per Joule than similarly sized dense models like Qwen3-32B and Yi-34B, while using substantially less peak VRAM and less energy per 1,000 generated tokens. The trade-off is a longer time-to-first-token because MoE routing adds prefill cost, and decode throughput gradually declines as context grows, dropping about 40 percent at 4K tokens compared with the 512-token baseline. Per active billion parameters, the model shows roughly an order-of-magnitude advantage in throughput and energy efficiency over the dense baselines, making it well suited to single-GPU production serving where memory and power budgets matter more than raw accuracy.

Eden AIdeepinfra/openai/gpt-oss-20bgpt-oss

Open provider page Provider docs

Quick Info

Provider: Eden AI
Model key: deepinfra/openai/gpt-oss-20b
Release date: Aug 5, 2025
Last updated: Aug 5, 2025
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.03
Output token cost: $0.14

Limits

Output tokens: 32,768 tokens
Context window: 131,072 tokens

Transparent token rates

Compare GPT OSS 20B (Deep Infra) pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about GPT OSS 20B (Deep Infra)

Eden AI

CoverageBenchmark

Inferbase's catalog page for OpenAI GPT-OSS-20B lists concrete technical specifications drawn from the model card: chat model type, 131,072-token context window for both input and output, 20.9B parameters, Apache-2.0 license, release date Aug 4 2025, training cutoff Jun 1 2024, and text-in/text-out modalities. A benchm The page attributes the model to OpenAI and provides an "OpenAI Docs" link for official documentation, but the supplied excerpt contains no mention of Deep Infra, Eden AI, or any inference-host/provider-specific behavior, so it is best read as creator-level documentation of the gpt-oss-20b family rather than news about

Eden AI

CoverageBenchmark

BenchmarkList publishes an evaluation page for OpenAI gpt-oss-20b that lists 114 benchmark rows and cross-compares it against named peers such as OpenAI Fable 5.1, Claude Opus 5, Kimi K3, Qwen3.8-Flash-Next, Qwen3.8-2.4T-A95B, and GLM 5.3 variants. The page surfaces a per-1M-token pricing row for gpt-oss-20b at $0.029 The excerpt does not name Deep Infra, Eden AI, or any hosting/inference provider, so this entry is best treated as a third-party cross-provider benchmark summary for the underlying OpenAI gpt-oss-20b model. Pricing and ranking shown are aggregator-level rather than Deep Infra or Eden AI specific, and the comparison set

Eden AI

CoverageAnalysis

An arXiv preprint (2508.16700v2) by Kumar, Yadav, and Patel presents a deployment-focused study of OpenAI's open-weight GPT-OSS-20B Mixture-of-Experts model on a single NVIDIA H100 GPU using bf16 precision. The paper reports that the model has approximately 20.9B total parameters with only about 3.61B active per token The study benchmarks GPT-OSS-20B against dense baselines Qwen3-32B and Yi-34B across time-to-first-token (TTFT), full-decode throughput (TPOT), end-to-end latency percentiles, peak VRAM with past key values held, and energy consumption via nvidia-smi. At a 2,048-token context with 64-token decode, GPT-OSS-20B delivers