NaiveAI releases open coding-model weights with a 315 GB footprint

Naive-N0.5-Flash has open weights and inference code under the MIT licence, with a stated million-token context window. Teams can examine and adapt the released material without waiting for a hosted service.

Access and ease of deployment are different questions. The repository specifies approximately 315 GB for the FP8 weights, plus additional memory for inference, and requires compatible NVIDIA hardware. Its mixture-of-experts design has 309 billion total parameters with 15.5 billion active for a token. The smaller active count describes computation; it does not turn the full model into a 15.5-billion-parameter download.

The release also describes a fast runtime called NaiveRT. That is separate from the inference code already present. The developer promises the runtime and benchmark material by October 12; API access is also forthcoming. We could inspect neither a working public API nor the promised runtime repository at this check.

Treat the speed headline as a vendor measurement. The technical article’s peak figure comes from a best one-second interval in a narrow generation test, excluding input processing. That cannot establish the elapsed time or correctness of a multi-step coding task.

A useful first evaluation therefore begins before downloading: estimate available memory, review the code and licence, then select a small set of your own long-context tasks with known acceptance criteria. Measure successful completion, latency and operating cost together. The opportunity is another controllable tool; whether it is a better tool for your work remains an empirical question.