~130K tokens. Supports 1M tokens and performs at 1M tokens are very different claims. Every architect should run their own needle-in-a-haystack tests — vendor benchmarks are not your workload.
nicornle