Can someone explain to me what I would need to do in terms of resources (GPU, I assume) if I want to run 20 concurrent processes, assuming I need 1k tokens/second throughput (on each, so 20 x 1k)
Also, is this model better/comparable for information extraction compared to gpt-4.1-nano, and would it be cheaper to host myself 20b?
Meta's goal with Llama was to target OpenAI with a "scorched earth" approach by releasing powerful open models to disrupt the competitive landscape. Looks like OpenAI is now using the same playbook.
I'm disappointed that the smallest model size is 21B parameters, which strongly restricts how it can be run on personal hardware. Most competitors have released a 3B/7B model for that purpose.
For self-hosting, it's smart that they targeted a 16GB VRAM config for it since that's the size of the most cost-effective server GPUs, but I suspect "native MXFP4 quantization" has quality caveats.
Please don't use the open-source term unless you ship the TBs of data downloaded from Anna's Archive that are required do build it yourself. And dont forget all the system prompts to censor the multiple topics that they don't want you to see.
Running a model comparable to o3 on a 24GB Mac Mini is absolutely wild. Seems like yesterday the idea of running frontier (at the time) models locally or on a mobile device was 5+ years out. At this rate, we'll be running such models in the next phone cycle.
Does anyone get the demos at https://www.gpt-oss.com to work, or are the servers down immediately after launch? I'm only getting the spinner after prompting.
The repeated safety testing delays might not be purely about technical risks like misuse or jailbreaks. Releasing open weights means relinquishing the control OpenAI has had since GPT-3. No rate limits, no enforceable RLHF guardrails, no audit trail. Unlike API access, open models can't be monitored or revoked. So safety may partly reflect OpenAI's internal reckoning with that irreversible shift in power, not just model alignment per se. What do you guys think?
Perhaps I missed it somewhere, but I find it frustrating that, unlike most other open weight models and despite this being an open release, OpenAI has chosen to provide pretty minimal transparency regarding model architecture and training. It's become the norm for LLama, Deepseek, Qwenn, Mistral and others to provide a pretty detailed write up on the model which allows researchers to advance and compare notes.
Can't wait to see third party benchmarks. The ones in the blog post are quite sparse and it doesn't seem possible to fully compare to other open models yet. But the few numbers available seem to suggest that this release will make all other non-multimodal open models obsolete.
What a day! Models aside, the Harmony Response Format[1] also seems pretty interesting and I wonder how much of an impact it might have in performance of these models.
149 comments
[ 2.3 ms ] story [ 89.0 ms ] threadWhat’s the catch?
Wonder if they feel the bar will be raised soon (GPT-5) and feel more comfortable releasing something this strong.
- OAI open source
- Opus 4.1
- Genie 3
- ElevenLabs Music
From a strategic perspective, I can't think of any reason they'd release this unless they were about to announce something which totally eclipses it?
so, the 20b model.
Can someone explain to me what I would need to do in terms of resources (GPU, I assume) if I want to run 20 concurrent processes, assuming I need 1k tokens/second throughput (on each, so 20 x 1k)
Also, is this model better/comparable for information extraction compared to gpt-4.1-nano, and would it be cheaper to host myself 20b?
For self-hosting, it's smart that they targeted a 16GB VRAM config for it since that's the size of the most cost-effective server GPUs, but I suspect "native MXFP4 quantization" has quality caveats.
Humanity’s Last Exam: gpt-oss-120b (tools): 19.0%, gpt-oss-120b (no tools): 14.9%, Qwen3-235B-A22B-Thinking-2507: 18.2%
[1]: https://msty.ai
Perhaps I missed it somewhere, but I find it frustrating that, unlike most other open weight models and despite this being an open release, OpenAI has chosen to provide pretty minimal transparency regarding model architecture and training. It's become the norm for LLama, Deepseek, Qwenn, Mistral and others to provide a pretty detailed write up on the model which allows researchers to advance and compare notes.
$0.15M in / $0.6-0.75M out
edit: Now Cerebras too at 3,815 tps for $0.25M / $0.69M out.
Super excited to test these out.
The benchmarks from 20B are blowing away major >500b models. Insane.
On my hardware.
43 tokens/sec.
I got an error with flash attention turning on. Cant run it with flash attention?
31,000 context is max it will allow or model wont load.
no kv or v quantization.
[1] https://github.com/openai/harmony