gpt-oss-120b
OpenAI · United States · Apache 2.0
Commercial use: Yes — free for commercial use
Apache 2.0. Still current, still OpenAI's open-weight flagship, and still one of the few frontier-adjacent US models under a fully permissive licence.
A large, capable model under Apache 2.0 that only activates a small fraction of itself for each word, so it runs far faster than its size suggests.
Why it should matter to you
The catch is memory rather than speed: you must hold all of it in RAM even though only a slice is used per word. Around 65 GB, so a serious workstation or an Ultra-class Mac. Cheap to run, expensive to house — a trade worth understanding before you budget hardware.
How it connects to our work
This shape is now standard at the frontier, and it changes buying advice: for local inference, prioritise memory capacity over raw compute. A machine with a fast processor and 32 GB will not run this; a slower one with 96 GB will.
Ships natively in MXFP4, so the footprint stays around 60-65 GB regardless of further quantisation.
Advantages
- Only 5.1B active parameters per token — very fast for its capability
- Apache 2.0 from a US lab, which matters where Chinese-origin weights face procurement resistance
- Reasoning effort is a per-request setting (low/medium/high) rather than a separate model
Disadvantages
- Needs ~65 GB, so an 80 GB GPU or 96 GB+ Mac
- Heavily safety-tuned and refuses more than most open models, which is a real friction in legitimate domains like security research and clinical text
- World knowledge is thinner than its total size suggests
Reach for it when
US-jurisdiction self-hosting where an Apache 2.0 licence and a Western lab are both requirements, and for agent loops where the reasoning dial saves real money.
Where it falls down
Domains that trip its safety tuning — medical, legal and security text all draw refusals that a less-tuned model handles. Test on your actual content before committing.
Running it
80 GB GPU, or 96 GB+ unified memory. See the hardware sizing tables for how that maps to specific chips and cards, and the quantisation guide for what you give up at each bit width.
Which cloud instance actually runs this, and what it costs
At Q4_K_M the weights are about 65 GB, so 10 catalogued instances fit with room for context, of which the 8 best value are shown. It is a mixture-of-experts model: all 65 GB must sit in memory, but each token reads only about 2.9 GB of active experts. The speeds below assume every weight is read, so treat them as a conservative floor and the costs as a ceiling. Measured single-user speeds for this kind of model run several times higher: about 100 tok/s for gpt-oss-120b on one A100 under vLLM, and around 240 tok/s for Qwen3.6 35B-A3B with multi-token prediction on an RTX 6000 GPU in Unsloth's tests. Cheapest per token is RunPod A100 PCIe (~$11.10 per million output tokens at an estimated 30 tok/s). Fastest on a single card is Azure NCads H100 v5 at roughly 52 tok/s for $37.62 per million tokens, which is the speed-versus-cost trade in one line.
| Instance | GPU | VRAM | Bandwidth | Est. tok/s | $/hr | $/M tokens |
|---|---|---|---|---|---|---|
| RunPod A100 PCIe | A100 80GB | 80 GB | 1935 GB/s | ~30 | $1.190 | $11.10 |
| Hetzner GEX131 | RTX PRO 6000 Blackwell Max-Q | 96 GB | 1792 GB/s | ~28 | $1.420 | $14.31 |
| RunPod H100 PCIe | H100 PCIe | 80 GB | 2000 GB/s | ~31 | $1.990 | $17.97 |
| Lambda H100 PCIe | H100 PCIe | 80 GB | 2000 GB/s | ~31 | $3.290 | $29.70 |
| Azure NC24ads A100 v4 | A100 80GB | 80 GB | 2039 GB/s | ~31 | $3.673 | $32.52 |
| Azure NCads H100 v5 | H100 | 94 GB | 3350 GB/s | ~52 | $6.980 | $37.62 |
| AWS g6e.12xlarge | 4x L40S | 192 GB | 864 GB/s | ~13 | $10.490 | $219.22 |
| AWS p5.48xlarge | 8x H100 | 640 GB | 3350 GB/s | ~52 | $55.040 | $296.65 |
Token generation on a single request is memory-bandwidth-bound, not compute-bound: the GPU must read every active weight from memory for each token it emits. So the ceiling is roughly GPU memory bandwidth divided by the size of the active weights. A 16 GB model on a 300 GB/s L4 tops out near 19 tokens/sec; the same model on a 2039 GB/s A100 tops out near 127. Figures on model pages are that arithmetic, computed from the bandwidth and weight-size columns here. They are a ceiling for one request at batch size 1, before serving overhead, so treat them as an upper bound for comparing instances rather than a benchmark. Batching raises total throughput well above this and does not raise per-request speed.
For identical silicon, the hyperscalers charge roughly 2 to 5 times what the specialist GPU clouds charge. An A100 80GB is about $3.43/GPU-hr on AWS, $3.67 on Azure and $5.03 on Google Cloud, against $1.19-1.39 on RunPod and $1.99 on Lambda. That spread is the single largest cost lever in a self-hosted inference budget, and it is larger than any saving from picking a smaller model.
Three costs that are absent from every headline rate: egress, storage for the weights, and idle time. A 70B model at Q4 is tens of gigabytes that must be pulled to the instance on every cold start, and an endpoint billed hourly is billed while it sits idle waiting for a request. Rates read 2026-09-06; check each provider's live pricing on the directory hub.
Free tiers carrying this model
| Provider | Type | The catch | Live limits |
|---|---|---|---|
| Groq | Inference host | Free-plan limits are per-model and vary widely across the catalogue. Token-per-day caps bite before request caps on long prompts. | check → |
| OpenRouter | Aggregator | Free-pool membership changes without notice — a model you built on can stop being free, and the list above will drift. Data passes through OpenRouter and then the upstream provider, so check the privacy and routing settings before sending anything sensitive. | check → |
| Cerebras Inference | Inference host | Small model catalogue compared with the aggregators. You are choosing it for speed, not selection. | check → |
| NVIDIA NIM | Inference host | It is a credit grant, not a permanent free tier. When the credits run out you are on a paid plan or self-hosting the container. | check → |
| Cloudflare Workers AI | Edge inference | The Neuron accounting makes it hard to predict how many actual requests you get, because different models consume at very different rates. | check → |
Jurisdiction: United States
Best raw capability and the deepest tooling ecosystem. For EU personal data you are relying on a transfer framework rather than on data never leaving the bloc, so check whether your DPA and your customers accept that.
Watch for: Enterprise API tiers usually promise no training on your data; consumer tiers and free tiers frequently do not. The free tier is where this bites.
Frequently asked
Can I use gpt-oss-120b commercially?
Apache 2.0. Still current, still OpenAI's open-weight flagship, and still one of the few frontier-adjacent US models under a fully permissive licence.
What hardware do I need to run gpt-oss-120b?
80 GB GPU, or 96 GB+ unified memory. Weights alone are roughly 65 GB at Q4_K_M and 63 GB at Q8_0. Ships natively in MXFP4, so the footprint stays around 60-65 GB regardless of further quantisation.. Add KV cache on top of that, which grows with your context length.
What licence is gpt-oss-120b released under?
Apache 2.0. Full commercial use, modification and redistribution. Patent grant included. The most permissive licence in common use for open-weight models.
Where can I use gpt-oss-120b for free?
Free tiers carrying it include Groq, OpenRouter, Cerebras Inference, NVIDIA NIM, Cloudflare Workers AI. Limits differ per provider and change often, so check each provider's own limits page. You can also self-host the weights, which has no rate limit at all.
Similar models
gpt-oss-20b
16 GB GPU or Mac. A strong default for local agents.
DeepSeek-V4-Flash
Hosted, or 256 GB+ unified memory at Q4. Not a laptop model.
GLM-5.3-Flash
80-96 GB at Q4_K_M. A workstation or Ultra-class Mac.
Qwen3.8 27B
24 GB GPU at Q4_K_M, or a 32 GB Mac at Q6_K.
Wiring gpt-oss-120b into something real?
We build the evaluation harness, the failover and the cost ceilings around a model like this, so it survives contact with production.
Maps to AI agent development, AI automation development and AI strategy and consulting. Or see it working: our case studies.