HighVRAM’s new gpt-oss-120b hardware guide turns OpenAI’s published memory boundary into six practical deployment paths. The core evidence is specific: gpt-oss-120b needs at least 60GB of VRAM in the cited MXFP4 Transformers path, or a multi-GPU setup, and OpenAI says it fits on one H100.
What that guidance does—and does not—prove
| Path | Why it belongs in the decision | What still needs proof |
|---|---|---|
| 64GB | First common label above the ≥60GB floor | Only 4GB nominal margin before cache and runtime state |
| H100 80GB | OpenAI’s cited single-GPU class | Exact PCIe/SXM/NVL identity, host, context and concurrency |
| RTX PRO 6000 96GB | One large current-generation CUDA pool | MXFP4 kernel support, measured peak and whole-system requirements |
| DGX Spark / GB10 | 128GB shared memory in a compact Blackwell system | Arm64 runtime support, host sharing and measured throughput |
| Large-memory Mac | Capacity for compatible converted artifacts | A separate MLX/Metal/llama.cpp path—not the cited CUDA path |
| Multi-GPU | OpenAI documents tensor and expert parallelism | Placement, topology and each device’s peak allocation |
Three safeguards before buying
- Active parameters are not stored weights: 5.1B active describes MoE routing; gpt-oss-120b has 117B total parameters.
- Maximum context is not a fit guarantee: the 131,072-token capability still requires cache, buffers and workload-specific headroom.
- Multiple cards do not become one automatic pool: tensor/expert parallelism, per-device capacity and interconnect must be planned.
Follow the decision path
- Read the complete 120B hardware guide.
- Review the gpt-oss-20b versus 120b requirements.
- Compare 48GB through 96GB capacity paths and the GPU comparison directory.
- Use the VRAM calculator as a planning range, then validate in the intended runtime.
- Inspect the benchmark database or submit a controlled result.
- Stage the new private gpt-oss-120b watchlist preset before checking current hardware.
Published October 25, 2026. The ≥60GB and single-H100 statements are OpenAI guidance for the cited MXFP4 path. HighVRAM has not independently benchmarked every system. Inclusion is not endorsement or ranking.