Local Model Field Notes / 186 Studios

gpt-oss on disk: where size stops buying you anything

These two models are natively MXFP4, so the usual GGUF quant ladder barely moves the file size. The real decision is not F16 vs Q8 vs Q4. It is just 20B or 120B, and it is decided by the task and by what fits in your memory.

Sizes are exact byte counts from HuggingFace (bytes / 1e9), pulled 2026-08-02. Benchmarks are OpenAI's official model-card figures at high reasoning effort, pass@1.
gpt-oss-20B 21.5B / 3.6B active gpt-oss-120B 120.4B / 5.1B active
5.2× bigger on disk: 120B (63.4 GB) vs 20B (12.1 GB), same quant
2.6 GB total spread of the 120B quant ladder, Q4_K_M to F16. Quant hardly matters
+19 points 120B gains on HealthBench Hard, its widest lead over 20B
+0.8 points 120B gains on AIME 2025. On hard math they are a tie
01 / Download size

The size matrix

Every format for both models. Notice the bars inside each column are almost the same length: quantizing gpt-oss changes the total size by a rounding error, because roughly 90 percent of the weights are experts frozen at MXFP4 in every row.

Format gpt-oss-20B gpt-oss-120B
MXFP4 native release (safetensors) 13.76 GB
65.25 GB
as shipped
GGUF F16 “original precision” 13.79 GB
65.37 GB
skip
GGUF Q8_0 12.11 GB
63.39 GB
GGUF Q4_K_M smallest of the four 11.62 GB
62.77 GB
leanest
Bars are scaled within each model's own column (0 to that model's F16), so they show intra-model spread, not the 5.2× gap between models. The llama.cpp MXFP4 GGUF is not listed separately because it equals Q8_0 to within a few hundred bytes. Ranges across the full ladder (Q2_K to F16): 20B 11.5 to 13.8 GB, 120B 62.6 to 65.4 GB.

The one rule that saves you a download: skip F16 GGUF. A genuine full upcast would be about 43 GB (20B) and 241 GB (120B). The F16 GGUF is only 13.8 and 65.4 because llama.cpp keeps the experts in MXFP4 and upcasts only the small attention and embedding tensors. You pay ~2 GB extra for precision that was never in the weights to begin with.

02 / Why the ladder is flat

The experts never move

A naive F16 conversion would triple the file for zero quality gain. What actually ships stays small because the expert weights are locked at 4.25-bit MXFP4 no matter the label. Each pair below is on its own scale.

gpt-oss-20B · 21.5B params
F16 GGUF
as shipped
13.8 GB
Naive full upcast
hypothetical
~43 GB
gpt-oss-120B · 120.4B params
F16 GGUF
as shipped
65.4 GB
Naive full upcast
hypothetical
~241 GB
What actually ships (experts frozen at MXFP4) Hypothetical naive upcast

Practical read: for gpt-oss there is no meaningful size-vs-quality dial inside a model. Take native MXFP4 (or equivalently the MXFP4 / Q8_0 GGUF), or Q4_K_M if you want the last half-gig. Unsloth measured even a 2-bit build performing "nearly the same" as F16, because only the tiny non-expert slice is affected. The quant question is basically solved. The model-size question is the whole game.

03 / Quality

Where the 120B earns its 5×

Official scores, high reasoning effort, sorted by the size of 120B's lead. The advantage is concentrated in health, agentic tool use, and graduate-level knowledge. On saturated contest math and on SWE-bench, the small model is right there with it.

gpt-oss-20B gpt-oss-120B bars = % accuracy · right column = 120B lead
HealthBench Hard
10.8
30.0
+19.2
HealthBench
42.5
57.6
+15.1
Tau-Bench Retail agentic tool use
54.8
67.8
+13.0
GPQA Diamond PhD-level, no tools
71.5
80.1
+8.6
MMMLU multilingual
75.7
81.3
+5.6
MMLU
85.3
90.0
+4.7
Humanity's Last Exam no tools
10.9
14.9
+4.0
AIME 2024 no tools
92.1
95.8
+3.7
SWE-bench Verified real code fixes
60.7
62.4
+1.7
AIME 2025 no tools
91.7
92.5
+0.8
0255075100%
Also worth noting, not shown above because it uses a different unit: Codeforces Elo (no tools) is 2230 for 20B vs 2463 for 120B. OpenAI positions 120B near o4-mini and 20B near o3-mini.
04 / The third lever

Reasoning effort moves more than model size

Before you reach for the bigger model, know that gpt-oss has a low / medium / high effort dial, and it swings scores harder than the 20B-to-120B jump does. Here is 120B on AIME 2025, no tools. Many chat apps default to low.

50
low
80
medium
92
high
A 42-point swing from one setting. If a local model feels dumb out of the box, check the effort setting before you blame the size.
05 / Crossover

So which one, and when

Since quant is a non-decision, the crossover is purely 20B vs 120B, governed by the task and by memory. Here is the split, then how it lands on your actual machine.

Reach for 20B MXFP4 · ~12 GB

  • Speed matters: it fits entirely in a 16 GB GPU and runs many times faster than a RAM-offloaded 120B
  • Contest-style math and general chat, where it scores within a point of 120B
  • Agentic loops that make many model calls, where latency compounds
  • Coding assistance: SWE-bench is 60.7 vs 62.4, effectively the same

Reach for 120B MXFP4/Q4_K_M · ~63 GB

  • Health, safety, or high-stakes reasoning: the HealthBench and GPQA gaps are large and real
  • Complex agentic tool use: Tau-Bench is +13 points
  • Broad or multilingual world knowledge where breadth counts
  • You have the memory to hold ~63 GB and can accept single-digit-to-low-tens tok/s
On your box: RTX 5080 (16 GB VRAM) + 64 GB RAM

The bars show each model's weight footprint against what your hardware can hold. The 20B lives on the GPU. The 120B only works because your GPU absorbs ~16 GB of it, leaving ~47 GB for system RAM instead of the full 63.

gpt-oss-20B Q4_K_M, 11.6 GBfits in VRAM, fully on the GPU
11.6 GB
scale: 0 to your 16 GB VRAM
gpt-oss-120B Q4_K_M, 62.8 GBGPU + system RAM, split
~16 on GPU
~47 GB in RAM
64 GB total
scale: 0 to your 80 GB combined (16 VRAM + 64 RAM). The GPU absorbs ~16 GB, so only ~47 GB lands in system RAM, leaving roughly 17 GB free for context and the OS

Verdict for your machine: run 20B MXFP4 as the fast daily driver, and keep 120B Q4_K_M for the hard health, knowledge, and tool-use questions where +13 to +19 points is worth dropping to a slower speed. Take Q4_K_M for the 120B specifically, because its 0.6 GB saving over Q8_0 is free headroom for your KV cache on a 64 GB box. If you find yourself living in the 120B, the upgrade that matters is more system RAM, not a bigger GPU.


Sources and caveats. Sizes: exact bytes from HuggingFace repos openai/, ggml-org/, and unsloth/ gpt-oss GGUF, verified against the HF API blobs endpoint. Benchmarks: OpenAI gpt-oss model card (arXiv 2508.10925), high reasoning effort, pass@1, no-tools variant where a model reports both. Param counts are HuggingFace's exact safetensors totals (21.5B / 120.4B); OpenAI's marketing rounds these to 21B / 117B. The naive-upcast figures (~43 / ~241 GB) are total-params × 2 bytes, a stated hypothetical, not a real download. The 2-bit quality note is from Unsloth and is community-sourced.