gpt-oss-120b

BaseRT .base builds of openai/gpt-oss-120b for fast local inference on Apple Silicon (Metal) and NVIDIA (CUDA).

A 117B-parameter mixture-of-experts reasoning model (36 layers, 128 experts, 4 active per token; ~5.1B active params) with alternating sliding-window / full attention, learned attention sinks and the harmony chat format. The checkpoint ships its expert weights in MXFP4; these builds carry those bytes verbatim — packed FP4 codes and E8M0 block scales, group 32, never dequantized or requantized — so the expert numerics are exactly the checkpoint's.

Files

File	Precision	Size
gpt-oss-120b-MXFP4.base	MXFP4 experts, bf16 attention / embeddings / lm_head	65.3 GB
gpt-oss-120b-Q8.base	MXFP4 experts, 8-bit attention projections, bf16 embeddings / lm_head	64.4 GB
gpt-oss-120b-Q4.base	MXFP4 experts, 4-bit attention projections, bf16 embeddings / lm_head	64.0 GB

MXFP4 is the checkpoint as published: the bf16 attention half untouched, so it is the reference build and loads on both backends. Q8 and Q4 narrow the attention projections to 8- and 4-bit (RTN from the bf16 weights, never quant-from-quant); the MXFP4 expert stacks, embeddings, lm_head, router, norms and biases are identical across all three, so the files differ only in the attention weights. A bare basert pull fetches Q4; ask for the others by width. All three need a machine with well over 64 GB of unified memory (a 96 GB or larger Mac, or a 128 GB GB10).

Usage

curl -LsSf https://basecompute.co/install.sh | sh
basert pull basecompute/gpt-oss-120b          # Q4
basert pull basecompute/gpt-oss-120b:q8       # Q8
basert pull basecompute/gpt-oss-120b:bf16     # MXFP4, as published
basert chat basecompute/gpt-oss-120b

Released under the apache-2.0 license, inherited from the base model.