GPT OSS 120B (Groq)

The luismori.dev engineering article 'gpt-oss-120b Speed, Benchmarks, and Routing Tradeoffs (2026)' describes gpt-oss-120b as a published OpenAI open-weight model released August 5, 2025 under Apache 2.0, with 117B total parameters, 5.1B active per token, 128K context, native MXFP4 quantization, and a single 80GB GPU d The piece frames gpt-oss-120b as attractive for engineers needing open weights, local or private deployment, provider choice, and fast decode, while recommending evaluation rather than default adoption for high-risk agentic work. It positions Groq alongside Cerebras, SambaNova, Fireworks, Baseten, Nebius, Google Vertex

The Markaicode page 'Groq GPT-OSS 120B Benchmark: Real Latency Methodology 2026' documents Groq's actual flagship production model, openai/gpt-oss-120b, describing it as a Mixture-of-Experts model with 120B total parameters and 5.1B active parameters per forward pass. The article cites Groq's production-models page fig The piece emphasizes Groq's hardware and precision angle — the LPU's TruePoint Numerics precision reduction and on-chip SRAM architecture — and breaks down latency into TTFT (queueing + prefill), decoding time (output tokens / generation speed), and network round trip. It also flags that 70B-and-above dense models see

The Markaicode guide 'Groq LPU Benchmark 2026' explicitly covers openai/gpt-oss-120b as the flagship general-purpose model on GroqCloud and corrects the premise that Groq runs on H100 GPUs. The page states that gpt-oss-120b at high reasoning effort posts a time-to-first-token of 0.71s, an output speed of 477 tokens/sec The article adds context that Groq's catalog is a curated set of open-weight models rather than a platform for arbitrary fine-tunes, and reports that Groq deprecated llama-3.3-70b-versatile for free/dev-tier traffic on June 17, 2026. Blended pricing across Groq's 11 tracked models ranges from $0.05/1M (Llama 3.1 8B) to

The seonova.pro SEO production benchmark explicitly lists 'GPT-OSS-120B (Groq)' as one of seven LLM models tested across 168 real SEO directory listings, assigning it a quality score of 7.2, a speed of 500 tokens/sec, and a per-listing cost noted as 'free (with rate limits that break production)'. The author recommends The article is grounded in eight months of production orchestration across 15+ models in real (non-synthetic) SEO content workflows. It documents the prompt template, evaluator methodology, and niche coverage (dentist, mechanic, Japanese restaurant, gym, traffic lawyer, hair salon, plumber, accounting firm), giving dev