What OpenAI released

OpenAI released two open-weight models under the Apache 2.0 licence: gpt-oss-120b and gpt-oss-20b. Both use a mixture-of-experts (MoE) architecture, support full-parameter fine-tuning, and are available for download from HuggingFace. Key characteristics verified at release:

  • Licence: Apache 2.0 — commercial use, modification, and redistribution permitted. No copyleft requirement on derivative works.
  • Architecture: Mixture-of-experts (MoE), consistent with the broader trend toward MoE for inference efficiency at scale
  • Fine-tuning: Full-parameter fine-tuning supported; LoRA and QLoRA adapters also work with standard HuggingFace PEFT tooling
  • Download: Available on HuggingFace under the OpenAI organisation namespace
Apache 2.0 matters

Llama 4 uses the Llama Community Licence, which prohibits use in products with more than 700 million monthly active users and requires attribution. Apache 2.0 has no such restrictions — you can deploy gpt-oss-120b in any product, at any scale, without licence compliance overhead.

Benchmark performance

Published benchmark scores for gpt-oss-120b at release (per OpenAI's model card — treat as self-reported until third-party replication):

Benchmark gpt-oss-120b gpt-oss-20b Llama 4 Maverick Gemma 4 27B
MMLU-Pro 90.0% ~82% (est.) ~85.5% (reported) ~78% (reported)
GPQA Diamond 80.9% — ~69% (reported) —
SWE-bench Verified 62.4% ~48% (est.) ~54% (reported) —
HumanEval — — — —

gpt-oss-20b benchmark figures are estimates based on typical capability scaling from 120b; OpenAI did not publish a separate 20b model card at launch. All competitor figures are from their respective model cards — cross-benchmark comparisons should be treated cautiously given evaluation setup differences.

Self-reported benchmarks

These figures come from OpenAI's own model card, not independent evaluation. GPQA Diamond at 80.9% would be a frontier result — run your own domain-specific evals before making infrastructure decisions based on published benchmark scores.

Hardware requirements

Model Precision VRAM required Recommended setup Use case
gpt-oss-120b fp16 ~80 GB 1× A100 80GB or H100 80GB Production serving, highest quality
int8 ~40 GB 2× A100 40GB Production, cost-optimised
int4 ~20 GB 2× RTX 4090 Dev/test; quality degradation expected
gpt-oss-20b fp16 ~16 GB 1× A10G or RTX 4090 Production, lower cost, fine-tuning
int8 ~8 GB 1× RTX 3090 or A10 Edge, constrained environments

For Indian cloud deployments: AWS instances with A100 80GB (p4d.24xlarge, ~$32/hr) and H100 80GB (p5.48xlarge, ~$98/hr) are available in ap-south-1 (Mumbai). Azure NC A100 v4 instances are available in Central India. Google Cloud A100 instances are available in asia-south1 (Mumbai).

The data-sovereignty angle for India and the UK

The most commercially significant aspect of these releases for Indian and UK enterprise builders is not benchmark performance — it is the ability to run GPT-4-class models entirely within your own infrastructure.

For India, the Digital Personal Data Protection (DPDP) Act creates specific restrictions on the transfer of personal data outside India, particularly for sensitive categories including financial data, health data, and children's data. Products that process any of these data types cannot send raw data to OpenAI's API without either a data processing agreement and transfer mechanism, or — in some interpretations — explicit user consent for cross-border transfer. Self-hosting gpt-oss-120b eliminates this compliance question entirely.

For the UK, the ICO's guidance on international data transfers under UK GDPR requires either adequacy decisions, standard contractual clauses, or binding corporate rules for data transferred outside the UK. For NHS-adjacent healthcare applications or FCA-regulated financial services, the compliance overhead of API-based LLM usage is substantial. On-premises deployment removes the transfer question from the compliance surface entirely.

Builder perspective

"We spent six months negotiating a data processing agreement with OpenAI to run our legal document analysis product. If gpt-oss-120b had been available then, we would have self-hosted from day one. The DPA negotiation alone cost us more in legal fees than a year of A100 GPU rental."

— A Verified Builder · London, UK

Comparing gpt-oss to the open-weight landscape

Model Parameters Licence Architecture Full fine-tune?
gpt-oss-120b 120B Apache 2.0 MoE Yes
gpt-oss-20b 20B Apache 2.0 MoE Yes
Llama 4 Maverick ~400B total / 17B active Llama Community MoE Yes
Gemma 4 27B 27B Gemma ToU Dense Yes
Mistral Devstral 2 ~22B Apache 2.0 Dense Yes

For more on the open-weight model landscape, see our April 2026 open-weight model roundup and Llama 4 deployment guide on HuggingFace.

Which workloads to migrate first

Not all workloads benefit equally from migrating to open weights. The cases where gpt-oss makes the strongest argument:

  • Data-sensitive document processing — legal, financial, healthcare documents that cannot be sent to a third-party API
  • High-volume structured extraction — invoice parsing, form extraction, data normalisation at scale where API cost is the primary constraint
  • Domain-specific fine-tuning — vernacular Indian-language models, sector-specific terminology, institutional writing style. Full-parameter fine-tuning on gpt-oss-20b is the most practical entry point
  • Offline or air-gapped environments — defence, critical infrastructure, on-device deployment where API connectivity is not possible

Cases where staying on the OpenAI API is still the right call:

  • Workloads that need GPT-5 class capability beyond what gpt-oss-120b provides
  • Teams without MLOps capability to manage model serving infrastructure
  • Low-volume products where GPU rental cost exceeds API cost
  • Products that rely on OpenAI's multimodal capabilities (DALL-E, Whisper, TTS) alongside the LLM

For a detailed cost comparison of when self-hosting beats managed APIs, see our inference economics analysis.