OpenAI Releases an Open-Source AI Model
GPT-OSS Brings Open-Source AI Back in a Big Way
Introduction
In a move that shocked much of the AI community, OpenAI has released GPT-OSS, its first truly open-source language model. This wasn’t just a half-hearted gesture or a watered-down version of an older model. GPT-OSS comes in two powerful sizes — 120 billion and 20 billion parameters — and it rivals some of the best proprietary models in the world.
Source URL here (gpt-oss-120b and gpt-oss-20b push the frontier of open-weight reasoning models)
This article will walk you through:
A short history of OpenAI’s changing relationship with open source
What it means to release “model weights”
How models like Deepseek may have pressured this decision
What GPT-OSS can actually do
What kind of computer you need to run it
And what this means for the future of AI
A Quick History: OpenAI, from Open to Closed, and Back?
OpenAI was founded in 2015 with a bold promise: to make artificial intelligence research open and accessible to everyone. In fact, “open” is literally in their name. But when they began to see serious progress with models like GPT-3, GPT-4, and the newer GPT-4o, something changed. OpenAI stopped releasing full models to the public. Instead, they shared only access through an API — meaning you could use their models but never look under the hood.
Critics even started jokingly calling them “ClosedAI.” Why? According to OpenAI, they were concerned about safety. Giving away super powerful models could lead to misuse — deepfakes, misinformation, or worse. [NOTE] This is an important point here. Open Source means free to download and use. Anyone with a computer powerful enough can run this from their school cafeteria. Like Deepseek did last year (here), we are seeing an unprecedented adoption of powerful AI models recently. What is very concerning about Deepseek, which is a Chinese model, is it does not have regular guardrails the way that the US-based models do. This is the most likely reason why OSS took so long to come out. It is said that once a model, with its open weights are in the wild, you can’t regain control over it.
But while OpenAI went quiet on open source, other groups didn’t. Models like Meta’s LLaMA, Mistral, and especially Deepseek — a rising star in China — began showing that open-source AI could compete with the best. And for developers, researchers, and startups, open-source models are a goldmine: free, modifiable, and able to run offline. Importantly, they do not “call home” or send the conversation prompts to the cloud where you don’t exactly know what your data is being used for. Be careful with open-source models. Ultimately, though, that growing pressure — and the fear of falling behind — may have forced OpenAI to change its stance.
What Does “Open Source” and “Model Weights” Mean?
Before diving into what GPT-OSS can do, let’s break down a few terms.
➤ Open Source
If a software program is open source, that means anyone can:
See how it works
Change it to suit their needs
Share improvements with others
In the world of AI, open source often includes both:
The code (how the model runs), and
The model weights — the actual “knowledge” the AI has learned during training.
➤ Model Weights
Think of a model’s “weights” as its brain. After training on tons of data, an AI model ends up with billions of numbers — these are the weights. They determine how it responds to your questions, how it writes essays, or how it solves math problems. Without weights, the model is like a blank brain. When a company shares them, it’s like sharing the full personality, skills, and memory of the AI. By releasing GPT-OSS with open weights under the Apache 2.0 license, OpenAI is saying: “Here’s our trained brain. Use it, modify it, even sell products with it.”
Did Deepseek Force OpenAI’s Hand?
The timing is interesting. Just weeks before GPT-OSS launched, Deepseek — a Chinese research team — released a stunning series of open-source models with strong performance. Unlike most Western tech companies, Chinese groups have leaned heavily into open source. That means:
The playing field is leveling
The best tools aren’t locked behind paywalls
And most importantly, China was starting to lead in open-source AI
For OpenAI — an American company — that may have been a wake-up call. In fact, U.S. government strategies on AI development have recently emphasized the need to keep “democratic” countries in the lead. So OpenAI's move may not just be about competition, but global positioning too.
What GPT-OSS Can Do (And Why It’s Impressive)
Let’s get into the meat of it. GPT-OSS isn’t just another open-source model. It’s shockingly good — especially the 120 billion parameter version, which:
Matches or beats GPT-3.5 in many benchmarks. Just a little context here, this statement isn’t very impressive, actually. GPT3.5 in “dog years” is pretty ancient, here’s some detail on this model (here). But remember it was the GPT3.5 model that exploded people’s awareness of the progress models had made and all of a sudden, user adoption went through the roof.
Excels in reasoning, code generation, tool use, and healthcare-related tasks
Supports long context lengths (up to 128,000 tokens)
Meanwhile, the 20 billion parameter version is scaled down considerably. It can:
Run on devices with just 16 GB of memory. Note this is not referring to GPU memory, but CPU memory. A decent laptop purchased in the last 12-months would likely have enough. My dev laptop that I use for coding has 48GB of RAM.
Beat most open models its size on coding tasks and reasoning
Be used for local inference — no cloud required
Expert Commentary
“We’re excited to make this model, the result of billions of dollars of research, available to the world to get AI into the hands of the most people possible,” said OpenAI CEO Sam Altman about the release. “As part of this, we are quite hopeful that this release will enable new kinds of research and the creation of new kinds of products.” He emphasized that he is “excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit.” Full article (here).
Head-to-Head with Proprietary Models
OpenAI’s own benchmarks show just how competitive GPT-OSS really is.
Test GPT-OSS (120B) GPT-3.5 (O3) Codeforces (coding contest) 2622 2706 MMLU (reasoning test) 90.0 93.4 HealthBench 57.6 59.8 GPQA Diamond (PhD science) 80.1 83.3 Tool Use (Function calling) 67.8 70.4 That’s incredibly close — within just a few points. And these are open models. Some experts are already saying this is the best US-based open-source model ever released.
Here’s a quick list of things this model can do — and it’s worth noting that both versions support these capabilities:
The model can be taught how to use external tools — like looking things up on the web, calling APIs, or solving code problems.
It can explain its thought process step-by-step, especially in math, science, or complex decision-making.
On medical Q&A tasks, it performs nearly as well as GPT-3.5.
How Safe Is It? Open Models Come with Risks
Let’s be honest — releasing a powerful model with open weights means anyone can use it, including bad actors. OpenAI tried to prepare for this:
They removed dangerous training data (e.g. info on bioweapons, cybersecurity hacks)
They fine-tuned the model with special safeguards
They tested how easy it was to break the model (and it held up well)
Still, once the model is out there, it can’t be recalled. Anyone can download, copy, and modify it. That’s the double-edged sword of open source. But OpenAI is putting its trust in the global developer and research community. It’s also offering $500,000 in rewards for “red teamers” who can find safety flaws. To me, this is more like cheap liability insurance if they ever get sued. Here is a great article explaining AI Red-Teams and what they do (here).
Summary
My gut instinct is that OpenAI OSS is really late to the open-source community. It’s important to understand that AI competency is growing and access to better tools are accelerating the types of innovation that we will be seeing. I’m not thinking that it’s generally a good idea to put these models out in the wild like this, but it is better to use US-based models because their reputation and credibility is on the line. Remember that OpenAI is partnering on Project Stargate. In my opinion, this is a mostly symbolic and defensive move. But it does show just how much pressure Sam Altman is under to drive this into people’s hands.
#Maranatha,
YBIC,
Scott