Views
No views yet
openai/gpt-oss-120b. Opt quants are per-tensor mixes of the six base quants, selected by measured KLD — measurement.json included. Additional quantizations can be generated by anyone using the measure.py output I recorded and the base quants.| Quant | Size | AIME24 | AIME25 | GPQA-D |
|---|---|---|---|---|
openai/gpt-oss-120b — published | 61 GB | 95.8 | 92.5 | 80.1 |
| 4.8bpw-opt-h16 | 70 GB | 90.0 −5.8 | 86.7 −5.8 | 73.7 −6.4 |
| 4.3bpw-opt-h16 | 63 GB | 93.3 −2.5 | 86.7 −5.8 | 74.7 −5.4 |
| 3.7bpw-opt-h16 | 55 GB | 86.7 −9.1 | 80.0 −12.5 | 74.7 −5.4 |
| 2.6bpw-opt-h16 | 39 GB | 83.3 −12.5 | 73.3 −19.2 | 71.7 −8.4 |
| 2.3bpw-opt-h16 | 35 GB | 83.3 −12.5 | 70.0 −22.5 | 71.7 −8.4 |
Don't use the branches at or above 4 bpw. gpt-oss-120b ships natively in MXFP4 — the MoE experts are already 4-bit (~4.25 bpw with block scales), only attention/embeddings/lm_head are BF16, and the full checkpoint is 61 GB. Quantizing above that adds size without adding information: extra bits re-encode already-4-bit values more faithfully, they don't recover precision the source never had. The 4.3 and 4.8 branches are larger than the original and show no quality advantage over it. Use the original. The sub-4 bpw branches are the ones that do something the original can't.
reasoning_effort=high, 122,880-token budget, temperature 0.| Branch | Bitrate | Size |
|---|---|---|
4.8bpw-opt-h16 | 4.8 bpw | 70 GB |
4.3bpw-opt-h16 | 4.3 bpw | 63 GB |
3.7bpw-opt-h16 | 3.7 bpw | 55 GB |
2.6bpw-opt-h16 | 2.6 bpw | 39 GB |
2.3bpw-opt-h16 | 2.3 bpw | 35 GB |
| Branch | Bitrate | Size |
|---|---|---|
8.0bpw-h16 | 8.0 bpw | 116 GB |
6.0bpw-h16 | 6.0 bpw | 87 GB |
5.0bpw-h16 | 5.0 bpw | 73 GB |
4.0bpw-h16 | 4.0 bpw | 59 GB |
3.0bpw-h16 | 3.0 bpw | 45 GB |
2.0bpw-h16 | 2.0 bpw | 31 GB |