Models / gpt-ossgenerated from capability-matrix.json

gpt-oss

OpenAI's open-weight gpt-oss, 20B and 120B.

chatexperts: sparse, no-sharedsafetensors, GGUF

gpt-oss 20B

Get it

tight pages experts from disk on MacBook Pro, M1 Pro, 16 GBfits experts streamed to the card on RTX 2070 SUPER, 8 GBfits on Ryzen 7 3700X, CPU only
goinfer-chat pull gpt-oss-20b
repo
ggml-org/gpt-oss-20b-GGUF
file
gpt-oss-20b-MXFP4.gguf
quant
mxfp4
size
12.1 GB
sha256
27cd6c432c7672cb…

Good for, and what it needs

The 20-35B-class MoE this project actually validates and measures: too big to hold fully resident on an 8 GB GPU, which is the point — bring -moe-cache-experts or -stream-weights.

~12 GB if loaded fully resident (native MXFP4); on an 8 GB card use -moe-cache-experts (CUDA resident-core + cached-experts, measured working, not just eligible — see this family's description above)

Tool calling

not yet measured

Decisions · /v1/systemone

Label scoring, unmeasured.

What hasn't been shown

  • Tool calling hasn't been measured on gpt-oss 20B.
  • No speed measured on any machine.
  • No graded speed on the M1 Pro 16 GB.
  • No graded speed on the RTX 2070 SUPER.
  • No graded speed on the Ryzen 7 3700X.
  • Never measured on a machine like yours, so no speed is shown.

How sure we are

Against the released model
100.0% picks the same next token as HuggingFace
0.99843 closest the raw scores get at their worst position (cosine, 1.00000 is identical; bar starts at 0.95)

real-oracle 100.0%/0.99843 · what parity-gated means

Measured speed · decode, tokens per second

your pick · MacBook Pro, M1 Pro, 16 GBMetal
not measured
your pick · RTX 2070 SUPER, 8 GBCUDA
not measured
your pick · Ryzen 7 3700X, CPU onlyCPU
not measured

Ollama v0.32.5 at its defaults, same machine, same session. 128 tokens of context unless noted. Full method in benchmarks.

Architecture, for the curious
model_type
gpt_oss
design
softmax-GQA
experts
sparse, no-shared
attention window
interleave
QK-norm
no
RoPE
full
norm
RMSNorm, pre-norm
activation
SwiGLU
tied head
no
modality
text
GPU-resident
eligible
registry description: OpenAI gpt-oss 20b/120b sparse MoE: per-head attention sinks + clamped interleaved-SwiGLU + alternating sliding/full + YaRN (MXFP4 experts; GPU-resident on BOTH Metal and CUDA since 2026-08-31 — the CUDA half validated on the real 20B, resident on an 8 GB card via --moe-cache-experts)