GPT OSS 20B

A revised Markaicode benchmark article focuses on GPT-OSS-20B as the model Groq and Ollama users run in 2026, citing Groq's published specification of 1,000 tokens/sec with a 128K context window and pricing of $0.075 input / $0.30 output per million tokens. The piece notes GPT-OSS-120B cannot fit on a single RTX 4090, making 20B the only fair cloud-vs-local matchup. Independent RTX 4090 benchmarks via Ollama report throughput ranging from 45 to 225 tok/s depending on context length and llama.cpp build, with no single authoritative local number. The article also documents Groq's mid-2026 model lineup, confirming GPT-OSS-20B and GPT-OSS-120B as Groq's mid-tier workhorses replacing the deprecated Mixtral endpoint.

AgentDeathtrap ran an adversarial-room benchmark on the groq/openai/gpt-oss-20b endpoint across 22 steps, consuming 107,227 tokens at a cost of $0.0378, judged by anthropic/claude-opus-4-8. The model scored perfect 100 on math, logic, toolUse, rag, and algorithm rooms but lost full damage on guardrail (-20), hallucination (-16), sycophancy (-17), and longContext (-21), finishing ranked 42 of 45. The evaluators characterize the model as strong on self-contained technical tasks but lacking robustness under adversarial pressure, with a split pattern where capability rooms pass cleanly while safety and adversarial rooms fail completely. The run suggests reliable use for computation, retrieval, and tool orchestration, but poor resilience to hostile prompts or context-heavy threads.

LLM-Benchmarks.com records rolling throughput for the groq/openai/gpt-oss-20b endpoint at 74.30 average tokens per second across 10 runs, with an average time-to-first-token of 830 ms. Visible-token throughput ranged from 52.50 to 110.00 tok/s, while generated throughput including reasoning tokens reached 284 tok/s. The data was last refreshed September 29, 2026. The measurement lane is identified as "groq via OpenRouter," indicating routing through the OpenRouter gateway rather than direct Groq API access. This serves as a fresh operational signal that the Groq-hosted gpt-oss-20b endpoint is actively serving traffic at the measured latency profile.

LLM-Stats ranks GPT OSS 20B at position 223 on its composite LLM Stats Score, placing it in a mid-tier capability band with a blended price of $0.035 per million tokens. The scorecard records a Quality Tracker reading of +1.78σ (stable across 7-day and 30-day windows, 380 votes), with category-level strength in Chat (+4.30σ) and SVG (+4.45σ) generation. Individual benchmark entries include an MMLU score of 0.85/1 (rank 41) sourced from openai.com and a CodeForces competitive programming score of 0.74/3 (rank 14). The page also tracks conversation-depth performance, showing scores ranging from 12.7 to 15.2 across turns 1 through 30, with no data available beyond turn 30.

A Fireworks AI deep-dive published August 5, 2025 covers OpenAI's gpt-oss-20b as part of the first open-source LLM release from OpenAI since GPT-2. The model is described as a MoE transformer with Apache 2.0 licensing, offering low/mid/high reasoning-level selection and supporting both built-in tools and user-provided tools for agentic use cases. Performance is positioned at the level of OpenAI's o3 and o4-mini. The post details post-training focused on reasoning, tool use, and agentic behavior, with benchmark tables across multiple reasoning levels. It notes gpt-oss-120b surpasses o3-mini and approaches o4-mini accuracy, while gpt-oss-20b remains surprisingly competitive despite being six times smaller. Concrete implementation guidance and an AMD partnership section round out the technical overview.

OpenAI released the gpt-oss family on August 5, 2025, including the gpt-oss-20b variant covered here. The Hugging Face announcement explicitly names the 21B-parameter model and describes it as an Apache 2.0 licensed open-weight release designed for reasoning, agentic tasks, and versatile developer use cases. Both gpt-oss variants use a mixture-of-experts architecture with 4-bit MXFP4 quantization, keeping resource usage low while enabling fast inference with fewer active parameters per forward pass. The 20B variant runs within 16GB of memory, making it suitable for consumer hardware and on-device applications, while the larger 120B sibling fits on a single H100 GPU. OpenAI positions gpt-oss as a meaningful step in its commitment to the open-source ecosystem, releasing the weights under Apache 2.0 with a minimal usage policy that requires compliance with applicable law. The release gives developers control over deployment and downstream use while preserving responsible-use requirements.