gpt-oss

gpt-oss-120b

OpenAI's gpt-oss materials say the 120b model is 117B total parameters and 5.1B active.

Apache-2.0, plus the OpenAI usage policy · 117B total, 5.1B active · context 131,072

OpenAI documents a 128k context for gpt-oss. The published config uses 131,072 positions.

From the model card

Strengths

  • OpenAI's model card describes a text model with configurable reasoning effort and agentic tool use, under Apache-2.0.
  • The 120b model is a mixture of experts: 117B total, 5.1B active, 128 experts, 4 active.

Weaknesses

  • OpenAI says the model will not work correctly without the harmony response format.
  • The card's memory figure is for native MXFP4 (within 80 GB, one H100). Weight memory on this page follows the 117B total at Q4_K_M or Q8_0, which is a different quant.
  • No Frankenstein row yet. Text only. The OpenAI usage policy also applies.

Sources: gpt-oss model card · Introducing gpt-oss · 120b docs · GitHub

Memory at 8,192 tokens

Computed, not measured. Weights use llama.cpp’s whole-file averages (Q4_K_M = 4.8944 bits/weight, Q8_0 = 8.5008) on the total parameter count. KV is fp16 and does not depend on the quant. The formula, including the sliding-window and latent-KV cases, is on the methodology page.

Computed memory for gpt-oss-120b at 8192 context
Q4_K_MQ8_0
Weights66.66 GiB115.79 GiB
KV at 8,1920.29 GiB0.29 GiB
Total66.95 GiB Computed116.07 GiB Computed

36 layers, 8 KV heads, head dim 64. The config alternates sliding attention (window 128) and full attention, and it starts with sliding.

Vendor figure, separate from the table: OpenAI says the 120b model runs within 80 GB in native MXFP4, on a single H100. That is their figure, not a measurement from this lab, and not the Q4_K_M total below. Source.

Tokens per second

Not yet measured. This lab has no Frankenstein row for gpt-oss-120b, so there is no estimated tok/s either.

Hardware tiers

Minimum is the smallest RAM that holds Q4_K_M plus 8k KV. Suggested does not name a second machine or a speed this lab has not run. Performance means a CUDA-class card with 12 GB or more of GDDR6 or GDDR6X and the weights fully on the card. The Tesla P40 is 24 GB of GDDR5, so a P40 speed is not the performance-tier number.

TierExample hardwareQuantFittok/s
MinimumCPU-only, with at least 68.95 GiB of system RAM for the Q4_K_M weights, 8k KV, and a 2 GiB planning allowance.Q4_K_MCPU-only, if the RAM is thereCPU: not yet measured
SuggestedA machine you are willing to wait on, or a GPU that can hold the quant. This lab has not measured a separate “comfortable” machine, so this row does not invent one.Q4_K_M until a Q8 run existsWhichever of the machines above actually holds the weightsnot yet measured
PerformanceA CUDA-class card with enough GDDR6 or GDDR6X to hold about 67.95 GiB (Q4_K_M weights, 8k KV, and a 1 GiB planning allowance). 12 GB is not enough for that total.Q4_K_MFull VRAM only if the card is larger than 12 GBnot yet measured

Where it sits

Best-for is a judgment from the card’s stated jobs and the model’s size. It is not a speed ranking. The matching build is CPU-only used Xeon. Score it against a machine you have on What can my rig run?