GPT-OSS 120B Hardware Requirements

Direct answer

Direct answer guidance: Choose Spark when you want a compact 128 GB appliance and context headroom; choose an 80–96 GB high-bandwidth GPU when throughput is the top priority.

Model / quant variants

Model / quant variants guidance: Use the officially supported deployment recipes and a current vLLM/TensorRT/llama.cpp implementation with MXFP4 support. Do not silently substitute a different quantization when comparing results.

Approximate weight footprint

Approximate weight footprint guidance: OpenAI states that gpt-oss-120b fits within 80 GB in MXFP4. Reserve additional memory for cache and runtime; Spark's 128 GB makes that easy, while exactly 80 GB needs tighter context planning.

Hardware tiers

Hardware tiers guidance: OpenAI states that gpt-oss-120b fits within 80 GB in MXFP4. Reserve additional memory for cache and runtime; Spark's 128 GB makes that easy, while exactly 80 GB needs tighter context planning. gpt-oss-20b fits a 16 GB-class device; Mistral Small 4 offers multimodal MoE capability; Qwen3.6 offers much smaller/faster coding-agent deployment.

Observed configurations

Observed configurations guidance: A public DGX Spark llama.cpp example above 50 tok/s shows strong interactive potential for a 120B model. Higher-bandwidth professional GPUs may improve decode and batching.

Context / concurrency effects

Context / concurrency effects guidance: Start with 16K–32K and a bounded concurrency. The extra memory should first protect reliability, then be allocated to longer context.

Performance expectations

Performance expectations guidance: A public DGX Spark llama.cpp example above 50 tok/s shows strong interactive potential for a 120B model. Higher-bandwidth professional GPUs may improve decode and batching.

Best-value configurations

Best-value configurations guidance: This is one of the strongest reasons to buy an 80–128 GB system because the model explicitly targets that memory class. Choose Spark when you want a compact 128 GB appliance and context headroom; choose an 80–96 GB high-bandwidth GPU when throughput is the top priority.

What will not fit / weak evidence

• Confirm the checkpoint format and runtime support rather than assuming every 4-bit conversion behaves like MXFP4.
• A reasoning model may generate more tokens per answer, so end-to-end latency matters alongside tok/s.

Planner CTA

Compare this recommendation against your model, context, concurrency, latency, and budget in the ComputeSage Planner.

Sources / methodology

Recommendations combine official specifications, public model/runtime documentation, adjacent benchmark observations, and clearly labeled engineering estimates. Review the ComputeSage evidence methodology.

• OpenAI gpt-oss introduction and deployment guidance
• NVIDIA DGX Spark specifications
• NVIDIA RTX PRO 6000 Blackwell specifications
• vLLM installation and hardware documentation
• llama.cpp project and backend documentation
• ggml-org llama.cpp DGX Spark benchmark