OpenAI's First Major Open-Weight Release, One Cycle In
OpenAI released GPT-OSS (its first significant open-weight family) in two variants, 20B and 120B parameters, in mid-2025. The release represented OpenAI's strategic pivot toward open-weight participation after years of closed-API-only deployment. By May 2026, the family has accumulated meaningful adoption metrics. This page tracks where GPT-OSS stands against the dominant open-weight families (Llama, Qwen, DeepSeek).
Headline Adoption Metrics (May 2026)
| Metric | gpt-oss-20b | gpt-oss-120b |
|---|---|---|
| Hugging Face Downloads (all-time) | ~7.3 million | ~4.6 million |
| Hugging Face Likes | 4,604 | 4,772 |
| Like-to-Download Ratio | ~0.06% | ~0.10% (very high) |
| Hosted Endpoints (Groq, Together, Fireworks, etc.) | 20+ | 10+ |
| Groq Pricing | $0.075 / $0.30 per 1M (in/out) | $0.15 / $0.60 per 1M (in/out) |
| Inference Speed (Groq LPU) | ~1,000 tps | ~500 tps |
Position vs Major Open-Weight Families
| Family | Top Variant Downloads (HF, all-time) | Top-20 Text-Gen Position |
|---|---|---|
| Qwen (Alibaba) | Qwen 3.5 0.6B: 19.0M | 11 of top 20 |
| Llama (Meta) | Llama 3.1-8B-Instruct: 9.8M | 3 of top 20 |
| GPT-OSS (OpenAI) | gpt-oss-20b: 7.3M | 2 of top 20 |
| DeepSeek | DeepSeek V3.2: 4.1M | 2 of top 20 |
| Mistral | Mistral-7B-Instruct-v0.2: 3.2M | 1 of top 20 |
Six Things the Adoption Data Tells You
- GPT-OSS placed in the open-weight top 20 within a year. Combined ~11.9 million downloads is enough to make GPT-OSS the 4th-largest open-weight family on Hugging Face by deployment, behind only Qwen, Llama, and the legacy giants (GPT-2, OPT-125m). Strong debut for a brand-new entrant.
- The 120B variant has the highest like-to-download ratio in the cohort. 0.10 percent versus 0.06 percent for gpt-oss-20b and ~0.04 percent typical. The pattern reflects strong quality perception among the developers who chose to deploy it.
- Groq is the leading inference partner. gpt-oss-20b at ~1,000 tps and gpt-oss-120b at ~500 tps on Groq's LPU hardware. Pricing is competitive with the cost-leader tier (Llama 3.1-8B at $0.05 / $0.08). Together, Fireworks, and others host both variants.
- Qwen still dominates open-weight deployment. 11 of 20 most-downloaded text-generation models on HF are Qwen variants; Llama is at 3, GPT-OSS at 2, DeepSeek at 2. The "Chinese open-weight families lead deployment" narrative is structurally robust regardless of OpenAI's entry.
- GPT-OSS is not threatening Llama's position. Llama 3.1-8B alone (9.8M downloads) exceeds the entire GPT-OSS family combined (11.9M across two variants). Meta's Llama 4 ecosystem reach (Meta AI consumer surface + Llama API + multi-hoster) substantially exceeds OpenAI's open-weight reach.
- The strategic value is partly defensive. OpenAI's GPT-OSS release was partly a hedge against developer concern about closed-API dependency and partly a response to DeepSeek V3's competitive cost-quality position. Treating GPT-OSS as a pure commercial-revenue play would understate the strategic signal.
What This Means for AI Visibility
GPT-OSS adoption matters for brand visibility in two specific ways. First, agent frameworks that route to multiple open-weight models now include GPT-OSS in their model menus; brand recall on GPT-OSS may differ from brand recall on Llama or Qwen due to OpenAI's distinct training-data composition. Second, GPT-OSS deployment is concentrated among self-hosting enterprise developers, which is a high-value demographic for B2B brand visibility. Brands tracking only the cloud-API surface miss the GPT-OSS self-hosted footprint.
Methodology
Hugging Face download and like data pulled from the HF Hub API on May 14, 2026. Hosted endpoint count from Groq, Together AI, Fireworks AI, Replicate, and other inference-provider directories. Comparative open-weight ranking from our companion Hugging Face Most-Downloaded Models page. Refreshed quarterly.
How Presenc AI Helps
Presenc AI tracks brand-mention rates across open-weight model deployments alongside cloud-API surfaces. When a brand performs well on GPT-OSS but underperforms on Llama or Qwen (or vice versa), the gap signals training-data composition differences that matter for downstream agent stacks. For brands with enterprise self-hosting exposure, this open-weight tracking is structural to total AI visibility.