Yesterday, OpenAI unveiled gpt-oss-120b and gpt-oss-20b, its first open-weight reasoning models since GPT‑2 back in 2019 and the AI world hasn’t stopped buzzing since.
These are not stripped-down or experimental releases. gpt-oss-120b matches OpenAI’s o4-mini on core benchmarks and even surpasses it in some domains. The smaller gpt-oss-20b performs at the level of o3-mini – a lightweight but capable reasoning model – while being designed to run on everyday developer machines.
For commercial real estate professionals, this is a milestone. This marks a strong step forward for firms that want to eventually deploy AI models entirely offline, with full control over their reasoning, customization, and data privacy.
What’s the News?
OpenAI has released two open-weight models under the Apache 2.0 license:
- gpt-oss-120b: A 117B-parameter model that matches o4-mini in performance and runs on an 80GB GPU or equivalent high-end workstation.
- gpt-oss-20b: A 21B-parameter model that performs similarly to o3-mini and runs locally on machines with just 16GB of RAM.
Key capabilities include:
- Adjustable reasoning levels (low, medium, high)
- Tool use (e.g., function calling, Python execution, web search)
- Compatibility with structured outputs and agentic workflows
- Local and edge deployment with no cloud requirement
This release is a major shift from OpenAI, giving developers access to near-frontier models they can run, fine-tune, and modify completely independently.
Why This Matters for CRE Agents Users
Your Digital Coworker runs in the cloud, using hosted models managed by companies like OpenAI, Anthropic, and Google. That’s ideal for simplicity, scale, and speed, but not for everyone.
1. Private, Secure On-Prem Deployment
CRE firms dealing with confidential deal data, investor reports, or proprietary models will eventually be able to run AI locally without data ever leaving their secure environment.
2. You Own the Model, You Control the Customization
Fine-tune custom versions of these open models on internal data and conventions without involving any third-party API. You own the weights, the fine-tuning, and the environment.
3. Offline-First Workflows
For users operating in air-gapped networks, regulated industries, or limited-connectivity regions, gpt-oss opens the door to full-featured, offline AI workstreams.
Let’s be clear: CRE Agents doesn’t yet offer local LLM deployment. But this release brings that reality closer and more feasible.
A Win for Control, Compliance, and Flexibility
gpt-oss doesn’t replace hosted models, but it expands the spectrum of choice.
- Use hosted models for seamless updates and fast integration
- Use local models for maximum privacy and control
- Combine both in a hybrid workflow as needed
The future of CRE AI doesn’t need to be tied to a proprietary cloud. Open weight models like gpt-oss show that to be true.
Interested in being one of the first to teach a digital coworker to work like you do?
We’re regularly opening up slots for early adopters to train and use digital coworkers with CRE Agents. Learn more and join the list to be next.
