TextMixture-of-Experts116.83B params5B active131K max context

GPT-OSS 120B (MoE) VRAM Requirements: How Many GPUs & What It Costs

GPT-OSS 120B (MoE) needs 64.3 GB of VRAM for its weights at AWQ 4-bit — 64.6 GB once a KV cache is loaded. Below is the full breakdown: which GPUs fit, how many you need, what it serves, and what it costs per million tokens. Every figure comes from tokcalc's open-source capacity model.

Weights (AWQ 4-bit)
64.3 GB
64.6 GB with KV cache
Fits on 1 GPU?
Yes
H100 SXM5 80GB or larger
Cheapest config
$0.004393
8× RTX 3080 10GB / 1M tokens
Fastest config
166400
tok/s on 2× Cerebras CS-3 (WSE-3)
VRAM requirement by quantization
Weight footprint = params × bytes per param. The “with KV cache” column adds one request at 4K context.
QuantizationWeightsWith KV cacheSmallest single GPU
FP16233.7 GB234.0 GBB300 SXM 288GB (288 GB)
FP8 (E4M3)116.8 GB117.1 GBIntel Gaudi 3 128GB (128 GB)
AWQ 4-bitrecommended64.3 GB64.6 GBH100 SXM5 80GB (80 GB)
Which GPU should you use?
Reference workload: 8K context, 8 concurrent users, continuous batching, AWQ 4-bit. Ranked by fitting first, then by tokens per dollar-hour.
GPUGPUsTotal VRAMDecode tok/sAggregate tok/sCost / 1M tokens
RTX 3080 10GB8×80 GB1054101171$0.004393
RTX 4080 16GB8×128 GB994.295447$0.006985
RTX 5090 32GB4×128 GB124259638$0.009316
RTX 3090 24GB4×96 GB649.027922$0.009949
RTX A4000 16GB8×128 GB621.259638$0.0112
RTX 4090 24GB4×96 GB698.933546$0.0132
T4 16GB8×128 GB443.725412$0.0175
V100 SXM2 32GB4×128 GB624.029952$0.0185
A10G 24GB4×96 GB416.019968$0.0200
TPU v6e Trillium (32GB)4×128 GB567.827256$0.0326
A100 40GB SXM42×80 GB537.312896$0.0388
L4 24GB4×96 GB208.09984$0.0445

RTX 3080 10GB has no NVLink, so 8-way tensor parallelism runs over PCIe. tokcalc applies a flat 0.85× multi-GPU efficiency factor for this, which is realistic for NVLink-equipped datacenter cards but optimistic for PCIe scale-out — treat the throughput and cost figures for this configuration as an upper bound, and measure before committing.

How concurrency changes with context length
On 8× RTX 3080 10GB at AWQ 4-bit. KV cache scales linearly with context, so long contexts cut the number of users you can serve from the same GPUs.
ContextConcurrent usersKV per requestFits?
4K101309.2 MB✓ fits
8K51618.5 MB✓ fits
33K122.42 GB✓ fits
66K64.83 GB✓ fits
131K39.66 GB✓ fits
Frequently asked questions

Can GPT-OSS 120B (MoE) run on a single GPU?

Yes. At AWQ 4-bit the model needs 64.6 GB of VRAM, so it fits on a single H100 SXM5 80GB or anything with more VRAM. At full FP16 it needs 234.0 GB, which fits on one B300 SXM 288GB or larger.

How many GPUs does GPT-OSS 120B (MoE) need?

At AWQ 4-bit with 8,192-token context and 8 concurrent users, the cheapest configuration that fits is 8× RTX 3080 10GB (80 GB total VRAM). $0.004393 per 1M output tokens.

How fast can GPT-OSS 120B (MoE) be served?

On 2× Cerebras CS-3 (WSE-3) at AWQ 4-bit, the aggregate throughput is 166400 tokens/sec with continuous batching. That hardware has no published rental price, so its cost per million tokens is not comparable.

How much VRAM does GPT-OSS 120B (MoE) need?

Weights alone: 233.7 GB at FP16, 64.3 GB at AWQ 4-bit. Add KV cache for your context length and concurrency — every concurrent request at 8,192 tokens costs extra VRAM, so budget headroom above the weight figure.

Should I run GPT-OSS 120B (MoE) in FP8 or FP16?

FP8 halves the weight footprint (234 GB → 117 GB) and is faster than FP16 on Hopper and Blackwell, where it has native kernels. FP16 remains the reference quality bar; FP8 loses a small amount of accuracy in exchange for roughly doubling the batch size you can serve per GPU.

Why does GPT-OSS 120B (MoE) need more VRAM than its active parameter count suggests?

GPT-OSS 120B (MoE) is a mixture-of-experts model: 5B parameters are active per token, but all 117B experts must stay resident in VRAM. Capacity is bounded by the resident set, while decode throughput is bounded by the active set — so it serves fast per token but needs the full footprint.

Is it cheaper to self-host GPT-OSS 120B (MoE) than to call an API?

At 8× RTX 3080 10GB the marginal cost is $0.004393 per 1M output tokens, before you add idle capacity, ops staff, or egress. That is typically cheaper than frontier API pricing at sustained volume, but API calls win when your load is bursty or too small to keep a GPU busy. The Build vs Buy tab on tokcalc prices this against a specific provider for you.

Try GPT-OSS 120B (MoE) with your own context and concurrency

The calculator uses the exact same formulas as this page — change the workload and watch every number update.

Open in calculator

GPT-OSS 120B (MoE) on specific GPUs

GPT-OSS 120B (MoE) on RTX 5090 32GBGPT-OSS 120B (MoE) on AMD MI300X 192GBGPT-OSS 120B (MoE) on L40S 48GBGPT-OSS 120B (MoE) on AMD MI325X 288GBGPT-OSS 120B (MoE) on A100 80GB SXM4GPT-OSS 120B (MoE) on Intel Gaudi 3 128GBGPT-OSS 120B (MoE) on H100 SXM5 80GBGPT-OSS 120B (MoE) on B200 SXM 192GB
Model architecture
The numbers behind the KV-cache and throughput math.
Total parameters116.83B
Active parameters per token5.10B
ArchitectureMixture-of-Experts
Attention layoutHybrid: 18 full + 18 sliding
Transformer layers36
Hidden dimension2880
Attention heads64 query / 8 KV (GQA)
Head dimension64
Vocabulary size201,088
Max context (catalog)131K
Native context4K
Context extensionRoPE scaling (yarn)
KV cache per token @ 8K36.6 KB
Native checkpoint size65.25 GB
Native weight formatmxfp4
Licenseapache-2.0

sliding layers stop growing once the context passes their window, so long-context cost is bounded.

Limitations of these numbers
Where the source does not fully establish the values above. Some of the figures hold only under conditions that may not apply to you.
  • •Active parameter count (5.1B) is taken from the model card; the checkpoint config does not state it. Weight bytes, parameter count and attention layout are exact.