OpenAI has finally released GPT OSS 20B!! This is probably the most powerful model you can run locally on a laptop. → Only 3.6B active parameters → Adjustable "Reasoning Effort" → Works with 16GB of RAM! Quick steps to use it 100% free and offline:

You can run it using LM Studio: 1. Download LM Studio for your OS (Windows, macOS or Linux) 2. In the search tab, type "GPT OSS 20B" 3. Download the recommended model

Then you're ready! Just open the chat tab and load the model from the top bar. Type your prompt, set the "Reasoning Effort", enjoy!

Useful links: Official blog post w/ benchmarks → This model runs really well on a laptop (+29 tokens/s on my config) but you can also try it for free online → Hugging Face →

been running it all day for reformatting and summarization tasks exactly as you said - 16GB is plenty and the adjustable reasoning is brilliant low effort for quick tasks = blazing fast high effort for analysis = deeper thinking switched everything from gpt-4 mini via groq because the speed difference is insane while quality stays solid finally have a capable model that doesnt need internet

Let's go, appreciate your feedback! I need to try the 120B version on Groq as well. The API price seems quite cheap.

in a day or 2 more inference providers will introduce this model and it should be even more affordable than groq (slower of course)

@itsPaulAi im curious what you use it for if you need a model that doesn't need internet and to be running it constantly? sry this is new to me and just genuinely curious

@itsPaulAi just took a look at your profile and looks like enterprise lol so ya def have zero clue what you would use it for since that's not my world 😂

@itsPaulAi even for personal, if you want to run your personal finance consultant, you dont want to put your bank details on an online model. lots of use cases.

why would you want to run this. I imagine it's extremely slow and just much better to use a pro sub. I understand you control your data, but this, on 16gb of ram, would be so slow

Total privacy + 100% free + runs offline + you can fine tune it, etc. Ofc it's less powerful than o3 but its speed is really good thanks to the MoE architecture.

I explained in this video – from what GPT-OSS is to how to run it locally with Ollama:

Almost killing my M1 Pro 16Gb. Maybe it is time to upgrade. What is the recommended Mac configuration to run this locally?

Good question, I'm not sure. I'm using a Snapdragon X Elite w/ 32GB RAM. So probably a M3 or M4 with the same amount of RAM?

i have 16 gb of ram and it doesnt work

Is this quantized ? It doesn’t run on 16 ram unless quantized

Yeah "MXFP4" by default... (first time I'm seeing this)

Its better than straight up 4 but quant but worse than 8 but

@itsPaulAi 😅just random daily hype cycle..quantized models are simply useless.

@NeuralAA Hard disagree here

@NeuralAA ok i hear you.. let me test it 😀

@NeuralAA Let me know how it goes :)

not running fast on 32 gb of my ram

If you've used it, your expectations should be this high

Just need a ram ? No need graphic card ?

@KevinMacFly1 Nope, if you're using shared (or better, unified) memory you're good. I'm only running it on my CPU and it's fast enough with 28+ tokens/s.

Because I have a ryzen 7 3700x + 32 gb ram with small gtx

Just tried it out. 20B model. Low, medium, high thinking. There should be an option to turn off thinking completely, just like qwen's /no-think

Been waiting for this! Curious how it compares to Llama 3 70B for complex reasoning tasks on consumer hardware.

• Surprisingly, fewer parameters yet enhanced reasoning capabilities suggest a focus on efficiency over sheer size in model design • This shift might indicate a trend towards optimized, localized AI processing

@grok what smarthpone can handle this?

Unfortunately at least the 20B one isn't very smart. Anything I threw at it, it got wrong.

Do you have any example? Will compare it to Qwen3 Coder 30B and Qwen3 30B Thinking.

A mother and his son are in a car crash. The mother dies, but the son is taken to the emergency room. At the OR, the surgeon looks at the patient and says: “I cannot operate on him. He’s my son.” How is this possible? Still thinks it's the mother with default reasoning.

Gave me this answer with medium thinking: Thought for 55.32 seconds Because the “surgeon” isn’t the boy’s biological mother – it’s his other parent. The man who says, “I can’t operate on him; he’s my son,” is the child’s father (or another parent who survived the crash). In short, the surgeon is the boy’s dad, not his mom, so the statement makes sense.

I'll try it with the high reasoning, it might help with the riddles but not sure if reasoning improves the knowledge of documentation.

The release of GPT OSS 20B by OpenAI is a game-changer for local AI processing. This move towards open-weight models not only enhances accessibility but also aligns with the growing demand for privacy and control in AI applications.

Devs are going to love this

Everyone honestly. You can write code ofc but also create agents, write 'regular' text, etc.

@TSEMAJ_41379

The corpus is an English‑heavy, historically‑back‑weighted web dump of many on‑line genres, with a long‑tail of topics and languages but a strong focus on high‑frequency words, high‑traffic domains, and an objective writing style. 1995–2020; 2021‑2022 🤔

Finally we can run serious models locally without melting our laptops. Privacy + performance is the dream combo.

Great test of GPT OSS 20B 👏👏👏

i have 16gb ram—model won't load.

Shared or unified memory? If so, it will load.

How is it compared to Deepseek R1? Interested in looking at the benchmarks

It’s running smoothly on a M3 pro with 36GB of RAM

Need a 8b version !!

@tweem 1.21x replies

Detailed introduction and user guide

Ram or vram? holy

Runs a little slow on a 4070 Super (12gb) but gets the job done. 100% and local baby, that's the way to do it.

Huge win for local AI.

Siri is cooked