MODELS
Open weights are becoming the default starting point
Open-weight models are now credible enough to benchmark first for many bounded workloads. The useful question is no longer open versus closed in the abstract, but which option clears the requirement at an acceptable cost and level of control.
What changed
OpenAI released the gpt-oss-120b and gpt-oss-20b model weights under the Apache 2.0 licence, with reference implementations for common local runtimes. OpenAI says the larger model can run in about 80 GB of memory and the smaller one in about 16 GB. That puts a reasoning model from a frontier lab inside hardware envelopes that an individual team can actually buy and operate.
The release is verified. The conclusion that open weights are becoming the default starting point is still developing: one release does not establish a market-wide default, and a model that fits in memory is not automatically the right model for the work.
Why it matters
The practical starting point is now a workload evaluation. Run a representative set of tasks against the smallest controllable model that might pass. Measure quality, latency, throughput, operating effort, and the cost of failure. Escalate to a hosted frontier model when the result requires it, not because the stack was chosen before the task was understood.
Open weights do not remove dependencies. A team still inherits model serving, updates, safety controls, observability, hardware capacity, and the licence terms of every component around the weights. Control becomes possible; it does not become automatic.
What we are watching
- Whether open models clear domain-specific evaluations before fine-tuning.
- The full cost of operating them, including engineering time and idle capacity.
- Hybrid designs that keep routine work local and reserve hosted models for the tasks that demonstrably need them.