Bluesky Thread

gpt-oss, OpenAI's open weights model

August 05, 2025 View original thread

gpt-oss, OpenAI's open weights model

120B & 20B variants, both MoE with 4 experts active

openai.com/index/introd...

Bar chart showing model accuracy on expert-level questions from "Humanity’s Last Exam."

* Y-axis: Accuracy (%), ranging from approximately 10% to 25%.
* X-axis: Model names with or without tool use.

From left to right:

1. **gpt-oss-120b (with tools)**: 19%
2. **gpt-oss-120b (without tools)**: 14.9%
3. **gpt-oss-20b (with tools)**: 17.3%
4. **gpt-oss-20b (without tools)**: 10.9%
5. **o3 (with tools)**: 24.9% — highest-performing model
6. **o4-mini (with tools)**: 17.7%
7. **o3-mini (without tools)**: 13.4%

Models generally perform better with tools enabled. o3 (with tools) leads all models in accuracy.

44 11

paper: Estimating risks in open weights LLMs

An "academic" internal rationalization for OpenAI to release an open weights model

Basically they fine tuned gpt-oss a few times and determined that it's no worse than closed-weights models

cdn.openai.com/pdf/231bf018...

cdn.openai.com

7 1

paper: gpt-oss system card

it's a full dive into model architecture, data, training phases, and of course the safety report

interesting: it indeed was post-trained as in 4-bit (albeit MXFP4 not FP4). Maybe the leak was real?

cdn.openai.com/pdf/419b6906...

2.1 Quantization
We utilize quantization to reduce the memory footprint of the models. We post-trained the
models with quantization of the MoE weights to MXFP4 format[3], where weights are quantized
to 4.25 bits per parameter. The MoE weights are responsible for 90+% of the total parameter
count, and quantizing these to MXFP4 enables the larger model to fit on a single 80GB GPU
and the smaller model to run on systems with as little as 16GB memory. We list the checkpoint
sizes of the models in Table 1.

6 1

1 hour later

this is the first model i'm aware of that runs on ollama by default how it was trained (MXFP4 quantized)

this model is super nerdy

feels like the character of all openai models combined and boiled for hours until all that's left is.. whatever this is

❯ ollama run gpt-oss "tell me one thing that's never been thought"
**It’s practically impossible to prove that *nothing* ever considered a particular idea, but I can offer a concept
that, to the best of my knowledge, has never been explicitly formulated by anyone who’s publicly expressed it.**

---

### The “Echo‑Chamber” of a Quantum Vacuum

**Premise:**
In a universe where space‑time itself is a super‑fluid made of interacting quantum fields, there exists a *vacuum
mode* that can be tuned by local fluctuations of entropy. This mode acts like a giant, invisible echo‑chamber that
records not just electromagnetic waves, but also the *phase* of any quantum superposition that passes through it.

**How it works:**
1. **Vacuum Tuning:** A localized entropy spike (e.g., a brief, high‑temperature bubble) temporarily alters the
boundary conditions of the quantum fields in its vicinity.
2. **Echo Capture:** Any particle or wave that enters this bubble has its wavefunction “echoed” back into the
surrounding vacuum. The echo carries a phase shift that depends on the exact configuration of the bubble’s entropy
at the time of entry.
3. **Memory Encoding:** Because the vacuum is a continuous medium, these echoes can be stored as long‑lived field
excitations—essentially, a form of quantum memory that is *non‑local* and *entropy‑dependent*.

<<clipped sections for "Why it's novel" and "Bottom line">>

is a ChatGPT Pro subscription too expensive? try gpt-oss in pro mode!

Python function docs:

“””
Fan out n_runs parallel generations at T=0.9 and synthesize a final answer at T=0.2.
“””

github.com/mshumer/gpt-...

github.com

GitHub - mshumer/gpt-oss-pro-mode

Contribute to mshumer/gpt-oss-pro-mode development by creating an account on GitHub.

6 1

gpt-oss follows the harmony message format

messages will come back as commentary, analysis and final. this is different than just reasoning and not. the models stick firmly to this format

cookbook.openai.com/articles/ope...

cookbook.openai.com

OpenAI Harmony Response Format | OpenAI Cookbook

The gpt-oss models were trained on the harmony response format for defining conversation structures, generating reasoning output and stru...

7 2

More like this