OpenAI · Local reasoning model · MXFP4
gpt-oss 20B VRAM and RAM requirements
Can you run gpt-oss 20B locally? OpenAI publishes a 16 GB memory minimum for its MXFP4 model. AIPCFit now uses it for a conservative baseline screen with explicit runtime, context, and confidence. Ollama lists a 14 GB download, which is a separate storage figure.
Sources verified . No hardware benchmarks are claimed.
Verified model specifications
- Model / publisher
- gpt-oss 20B / OpenAI
- Total parameters
- 21B (OpenAI rounded); 20.9B in Ollama
- Active parameters per token
- 3.6B
- Architecture
- 32 experts, 4 selected per token; 24 layers
- Context limit
- 131,072 tokens (128K)
- Weight quantization
- MXFP4 for MoE weights
- Official minimum memory
- 16 GB (published baseline; not guaranteed VRAM)
- Ollama download
- 14 GB, rounded listing size
- License
- Apache 2.0
The checkpoint configuration specifies 131,072 positions. Runtime defaults and practical memory limits may be lower. Sources: OpenAI architecture, checkpoint configuration, and Ollama artifact.
Total vs active MoE parameters
Mixture-of-experts routing selects part of the network for each token. The 3.6B active figure describes computation; it does not shrink the resident model to 3.6B parameters. Plan storage around the complete artifact and measure runtime memory separately.
MXFP4 is a floating-point quantization with block scaling. Ollama documents 4.25 bits per MoE weight and native support. Other tensors retain different precision, so multiplying every parameter by four bits cannot establish the download or runtime size. MXFP4 is not Q4_K_M. AIPCFit uses a separate MXFP4 planning representation; GGUF is recorded independently as the Ollama artifact container.
Download size vs required memory
The 14 GB figure comes from the published gpt-oss:20b package (manifest f38aa0c53da5). Allow disk space for the package and your other models. It is not a measured peak RAM or VRAM value.
Inference needs weights, KV cache, work buffers, and runtime allocations. Context length, concurrent requests, backend, and offload settings affect those allocations. The advertised 128K context is not a claim that 128K fits within the published 16 GB memory figure.
AIPCFit screens MXFP4 separately from GGUF heuristics using the published 16 GB memory baseline. Runtime and context allowances are unquantified by default. Baseline results have low confidence; missing runtime/context information, uncalibrated long context, and CPU/offload return Insufficient data.
NVIDIA GPUs: can gpt-oss 20B run on 16 GB VRAM?
This table assumes the verified Ollama MXFP4 artifact, a compatible backend, explicit 4096-token context, and resident GPU memory. The short-context limit is our screening policy, not the documented context of OpenAI's 16 GB claim. No numerical overhead is assumed.
| Memory budget | Conservative screen | Confidence |
|---|---|---|
| 8 GB | Below published baseline | medium |
| 12 GB | Below published baseline | medium |
| 16 GB | Tight — verify headroom | low |
| 24 GB | Baseline met — verify runtime | low |
8/12 GB are below the published baseline for this resident-memory screen; this does not prove CPU or offload execution impossible. 16 GB is tight. 24 GB meets the baseline with unquantified headroom. These are not benchmarked GPU compatibility claims.
Runtime/context overhead hooks accept sourced additional allowances above the baseline. No measured allowance ships with this release; without both allowances a longer context remains Insufficient data. The engine never adds system RAM to VRAM.
Treat a 16 GB NVIDIA configuration as a tight baseline candidate and test short context with free VRAM. The publisher's memory claim does not identify a universal driver, backend, or context combination. More VRAM leaves additional capacity, but capacity alone cannot verify MXFP4 execution.
Check Ollama's supported GPU and driver list. Its general GPU support is separate from a measured gpt-oss fit. The Transformers MXFP4 path has its own requirements, including compute capability 7.5 or later and compatible kernels/Triton. Do not transfer one runtime's requirements to another.
For smaller VRAM budgets, the below-baseline result applies to resident memory only; AIPCFit does not invent a CPU offload allocation. Check actual placement and memory use in your chosen runtime before relying on it.
Apple Silicon: gpt-oss 20B Mac RAM guidance
Ollama supports Apple GPU acceleration through Metal. Mac unified memory is shared by the operating system, applications, CPU, and GPU; total RAM is not an exclusive model budget.
Our planning inference is that a 16 GB Mac needs particular caution: it cannot reserve all 16 GB for inference while also running macOS. A 24 GB or 32 GB Mac has more total capacity, but these are not verified minimum or recommended tiers for gpt-oss. Check memory pressure with your actual runtime and context. AIPCFit screens safe available unified memory against the same baseline with low confidence; it does not certify a Mac fit or predict speed.
Runtimes, MXFP4 support, and Ollama installation
The official model card provides Ollama, Transformers, vLLM, PyTorch/Triton, and LM Studio paths. OpenAI also publishes a Metal reference implementation. Use a release that explicitly supports gpt-oss and its artifact format. This catalog entry records only the verified Ollama variant; support elsewhere does not mean the same package works unchanged.
Install the current Ollama release for your operating system, then run:
ollama pull gpt-oss:20b
ollama run gpt-oss:20b
ollama psInspect placement with ollama ps, start with short prompts, and monitor available memory. Use the runtime's Harmony chat template: the model card says the model will not work correctly without Harmony. Tool calling also needs an application that implements the requested tools.
Practical FAQs
Can I run gpt-oss 20B on 16 GB VRAM?
OpenAI publishes a 16 GB memory minimum for MXFP4. With a verified Ollama artifact and an explicit short context, AIPCFit marks a 16 GB budget as Tight with low confidence. A 24 GB budget meets the baseline but is not certified to fit. Runtime/backend allocations and available memory still need verification.
How much RAM does gpt-oss 20B need on a Mac?
The published 16 GB minimum is not a universal Mac total-RAM requirement. AIPCFit compares the available model-memory budget after macOS/apps, not total RAM, with this baseline. Apple Silicon memory is shared. Unknown runtimes, long context without calibrated allowances, and CPU offload return Insufficient data.
Is the 14 GB Ollama download the VRAM requirement?
No. It is the rounded package size listed for gpt-oss:20b. Runtime memory also includes the KV cache, work buffers, and backend overhead; it depends on context and settings.
Is gpt-oss 20B a 3.6B model or a Q4_K_M model?
Neither label describes this artifact. OpenAI reports 21B total parameters and 3.6B active per token. Its MoE weights use MXFP4. Active parameters describe routed computation, while the total weights still need storage. MXFP4 and Q4_K_M are different quantization schemes.
Plan and check your hardware
- VRAM calculator — choose runtime, context, and memory placement for the MXFP4 baseline screen.
- GPU catalog — compare recorded dedicated memory.
- Local model compatibility checker — compare your hardware with existing supported estimates.
- Compatibility tools and all local models.
Official references
Verified 2026-11-03. Hardware guidance above is conservative planning advice; no device-specific runtime measurements were performed.
