ALITEQ.

What it costs to run gpt-oss 120B (MoE)

Trying to decide what to buy? The best GPU for gpt-oss 120B (MoE)

120.4B paramsMoE · 128 experts, 4 active128k max contextapache-2.0

Memory needed, by quantisation

At 16k context. Weights + KV cache + ~0.8 GB overhead.

QuantWeightsKV cacheTotalQuality
F16224.3 GB0.8 GB225.9 GBFull precision. Reference quality, twice the size of Q8 for no practical gain in most chat use.
Q8_0119.2 GB0.8 GB120.7 GBEffectively lossless. Use when VRAM is not the constraint.
Q6_K92.0 GB0.8 GB93.5 GBVery close to Q8 at meaningfully less memory. A safe high-quality pick.
Q5_K_M79.6 GB0.8 GB81.2 GBSmall, hard-to-notice quality loss. Good balance.
Q5_077.7 GB0.8 GB79.2 GBOlder-style 5-bit. Q5_K_M is usually the better pick at the same size.
Q4_K_M68.0 GB0.8 GB69.6 GBThe community default. Best quality-per-gigabyte for most people.
Q4_063.8 GB0.8 GB65.4 GBOlder-style 4-bit, measurably worse than Q4_K_M at a similar size. Avoid unless required.
IQ4_XS59.6 GB0.8 GB61.1 GBNewer 4-bit, smaller than Q4_K_M with comparable quality. Needs a recent llama.cpp.
Q3_K_M54.8 GB0.8 GB56.4 GBNoticeable degradation. Use to fit a larger model that would otherwise not run.
Q2_K47.0 GB0.8 GB48.5 GBHeavy degradation. Almost always better to run a smaller model at Q4_K_M instead.
IQ3_XS46.3 GB0.8 GB47.8 GBAggressive. Usually better than a smaller model at Q4, but test before trusting it.

Cards that can run it

Best quantisation each card fits at 16k context, with live rental price where we track one.

GPUVRAMBest fitUses~tok/sRent from
Apple M3 Ultra512 GBF16226 GB
Apple M2 Ultra192 GBQ8_0121 GB
Apple M4 Max (16-core CPU / 40-core GPU)128 GBQ8_0tight121 GB
AMD Ryzen AI Max+ 395 (Strix Halo, 128 GB)128 GBQ8_0tight121 GB
NVIDIA DGX Spark (GB10 Grace Blackwell)128 GBQ8_0tight121 GB
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition96 GBQ6_Ktight94 GB$1.690/hr
NVIDIA RTX PRO 6000 Blackwell Workstation Edition96 GBQ6_Ktight94 GB
Apple M2 Max96 GBQ6_Ktight94 GB
Apple M4 Pro64 GBIQ4_XStight61 GB
NVIDIA RTX 6000 Ada Generation48 GBIQ3_XStight48 GB$0.356/hr
NVIDIA RTX A600048 GBIQ3_XStight48 GB$0.267/hr
AMD Radeon PRO W790048 GBIQ3_XStight48 GB

How these numbers are produced

  • Model shape comes from the model's own config.json — 36 layers, 64 attention heads, 8 KV heads.
  • Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
  • KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
  • Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
  • This is a mixture-of-experts model. Generation reads only the experts active for each token (4 of 128), so throughput is scored on active parameters, not the full 120B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Derived from architecture; runs ~15% above the vendor-published active count for this family (distinct GptOssForCausalLM architecture). Treat derived tok/s as optimistic.
  • Rental prices are pulled hourly from provider APIs (last updated Wed, 22 Jul 2026 22:20:08 GMT). Spot/interruptible pricing can change or vanish without notice.

Try it interactively

Drag the context slider and watch the KV cache fill the card

Worth reading before you buy

Other models