Skip to content

VRAM reference

gpt-oss-120b VRAM requirements

Fine-tuning gpt-oss-120b takes about 66 GB of GPU memory with QLoRA, 238 GB with LoRA on 16-bit weights and 1,873 GB for full fine-tuning, at 2,048 tokens per sequence. Running it takes 234 GB in 16-bit or 66 GB in 4-bit with a 8K-token context. It is a mixture of experts: 5.1B of its 116.8B parameters work on each token, but all of them must be in memory.

By Will Schulz · Architecture and prices checked 15 November 2026 · Prices as of November 2026

Fine-tuning memory

At 2,048 tokens per sequence, batch 1, LoRA rank 16, gradient checkpointing on
MethodWeightsTraining stateActivationsTotalSmallest single GPU80 GB GPUs needed
QLoRA (4-bit base)62 GB0.2 GB4.0 GB66 GBA100 80GB (80 GB)1
LoRA (16-bit base)234 GB0.2 GB4.0 GB238 GBNone4
Full fine-tuning234 GB1,636 GB4.0 GB1,873 GBNone27

On a 24 GB card such as the RTX 4090: QLoRA: no, it needs about 66 GB at 2K tokens, more than the 24 GB available.

Inference memory

One sequence, fp16 KV cache
PrecisionWeightsKV cache at 8KTotalAt 128K contextSmallest single GPU
16-bit (BF16/FP16)234 GB0.3 GB234 GB238 GBNone
8-bit117 GB0.3 GB117 GB122 GBH200 (141 GB)
4-bit65 GB0.3 GB66 GB70 GBA100 80GB (80 GB)

Which GPUs fit gpt-oss-120b

One GPU unless a count is shown. Follow a GPU's name for its maximum sequence length, batch size, LoRA rank and rental cost.

GPUQLoRALoRAFull fine-tuningRun, fp16Run, int8Run, int4
RTX 3090 24 GB4×12×87×11×6×4×
RTX 4090 24 GB4×12×87×11×6×4×
RTX 5090 32 GB3×9×66×9×5×3×
L40S 48 GB2×6×44×6×3×2×
A100 40GB 40 GB2×7×53×7×4×2×
A100 80GB 80 GBYes4×27×4×2×Yes
H100 SXM 80 GBYes4×27×4×2×Yes
H200 141 GBYes2×15×2×YesYes
B200 180 GBYes2×12×2×YesYes

What's specific to gpt-oss-120b

  • Ships quantised: MXFP4 expert weights make a 65.3 GB checkpoint, which is how it runs on a single 80 GB GPU.
  • 128 experts with 4 active per token: 5.1B active of 116.8B total, so it is fast for its size but memory-hungry.
  • Upcasting to 16-bit for LoRA needs about 234 GB of weights, so single-GPU fine-tuning means keeping the weights in 4-bit, which just fits an 80 GB H100 at short sequences.
Architecturegpt-oss-120b
Parameters116.8B total, 5.1B active per token
Experts128 per layer, 4 used per token
Layers36
Hidden size2,880
Expert MLP width2,880
Attention heads / KV heads64 / 8 (head dim 64)
Vocabulary201,088 tokens
Context128K tokens; 18 of 36 layers use a 128-token sliding window
LicenseApache 2.0
ReleasedAugust 2025

Try other settings

Pre-filled for gpt-oss-120b. Change the sequence length, batch, LoRA rank, context or GPU.

JobMemoryRTX 4090 (24 GB)Smallest single GPU
QLoRA (4-bit base)2K tokens × batch 1, rank 1666 GB4 GPUsA100 80GB (80 GB)
LoRA (16-bit base)2K tokens × batch 1, rank 16238 GB12 GPUsNone: needs several
Full fine-tuning2K tokens × batch 11,873 GB87 GPUsNone: needs several
Inference, 16-bit (BF16/FP16)8K-token context × 1 sequence234 GB11 GPUsNone: needs several
Inference, 8-bit8K-token context × 1 sequence117 GB6 GPUsH200 (141 GB)
Inference, 4-bit8K-token context × 1 sequence66 GB4 GPUsA100 80GB (80 GB)

A job fits when the estimate is within 90% of the card's memory. Training assumes gradient checkpointing and FlashAttention; multi-GPU counts assume the memory splits evenly (FSDP, ZeRO-3 or tensor parallelism).

How we estimate it

Every figure comes from the model's published architecture (layers, hidden size, attention and KV heads, vocabulary) rather than a rule of thumb, using the same functions as the calculator above. Sizes are in GB (10⁹ bytes).

  • Full fine-tuning: 16 bytes per parameter for mixed-precision Adam (bf16 weights and gradients, fp32 master weights and two moments), as in the ZeRO paper.
  • LoRA: frozen bf16 weights (2 bytes per parameter) plus 16 bytes per adapter parameter, rank 16 on every linear layer. For mixture-of-experts models, adapters go on attention only and the experts stay frozen.
  • QLoRA: NF4 weights with double quantisation (4.127 bits per parameter); embeddings and the output head stay in bf16, as bitsandbytes leaves them.
  • Activations: gradient checkpointing and FlashAttention. Each layer's bf16 input is kept, one layer is recomputed at a time, and the output logits take 8 bytes per vocabulary entry per token (fp32 values and gradients). Chunked cross-entropy, as in Unsloth or Liger kernels, cuts that last term.
  • Inference: 2, 1 or about 0.56 bytes per weight (4-bit with group-wise scales), plus an fp16 KV cache of 2 × layers × KV heads × head dim × 2 bytes per token; sliding-window layers only cache their window.
  • Fits means the estimate is within 90% of the card's listed memory; the rest covers the CUDA context, kernels and fragmentation. Multi-GPU counts assume memory splits evenly.

Frequently asked questions

Can I fine-tune gpt-oss-120b on a 24 GB GPU?
On an RTX 3090 or 4090 (24 GB): QLoRA: no, it needs about 66 GB at 2K tokens, more than the 24 GB available. LoRA: no, it needs about 238 GB at 2K tokens, more than the 24 GB available. Estimates use gradient checkpointing and leave 10% of memory free.
How much VRAM do you need to run gpt-oss-120b?
About 234 GB in 16-bit and 66 GB in 4-bit, at 8K tokens of context. The smallest single GPU here for 4-bit is the A100 80GB (80 GB); 16-bit needs several GPUs.
How many GPUs does full fine-tuning gpt-oss-120b take?
About 1,873 GB with mixed-precision Adam at 2K tokens, so at least 27 × 80 GB H100s with the training state sharded (FSDP or ZeRO-3). LoRA needs 238 GB and QLoRA 66 GB.
Why does gpt-oss-120b need so much memory if only 5.1B parameters are active?
It is a mixture of experts: each token uses 4 of 128 experts per layer, but which ones changes token to token, so all 116.8B parameters stay in GPU memory. Speed follows the active parameters; memory follows the total. Our LoRA figures adapt attention only; adapting every expert at rank 16 adds about 21 GB of adapter training state.
How long a context can gpt-oss-120b handle on an 80 GB GPU?
Its KV cache costs 4.8 GB at the full 128K tokens (fp16 cache). On one 80 GB GPU: 16-bit weights don't fit, 128K tokens with 4-bit weights.

Terms used on this page: LoRA, QLoRA, KV cache, gradient checkpointing, quantization, mixture of experts. For any model size, use the fine-tuning VRAM calculator; for the method in depth, the Academy lesson Memory first: what hardware can train what.

Sources

  1. [1]Hugging Face. openai/gpt-oss-120b config.json
  2. [2]OpenAI. openai/gpt-oss-120b model card
  3. [3]OpenAI. gpt-oss-120b & gpt-oss-20b Model Card (arXiv:2508.10925)
  4. [4]Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
  5. [5]Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  6. [6]Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
  7. [7]Korthikanti et al., 2022. Reducing Activation Recomputation in Large Transformer Models (arXiv:2205.05198)
  8. [8]Chen et al., 2016. Training Deep Nets with Sublinear Memory Cost (arXiv:1604.06174)

Architecture read from config.json and parameter count from the Hugging Face safetensors metadata on 15 November 2026.

Find out if you need to fine-tune at all.

Kojen's prompt tuning and model search run for real on your own test cases, on hosted models. Measure the cheaper levers before you rent a GPU.

Start training free