The pitch: freeze Qwen3-8B, bolt on two small projection heads (state and action), train those heads contrastively with bidirectional InfoNCE on ~60M Nemotron Q&A pairs, ~30M hard negatives, and ~1M agentic trajectories. The result only scores candidates you hand it — no generation, probabilities are relative to the candidate set, and the encoder is locked to Qwen3-8B last-token pooling. The practitioner-relevant numbers: zero-shot it's claimed on par with Jev on computer-use, gaming, and tool-calling at up to 9× lower latency; fine-tuned heads hit 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1. Action embeddings are cached separately, so with ~1k candidates you get 13× over the baseline. Fine-tuning is cheap because only the heads train — the SOTA numbers come from those fine-tuned heads, not this checkpoint, which the card says plainly and deserves credit for. Azalia Mirhoseini on the author list, Apache 2.0 on everything, and a "multimodal CLM-35B coming in early October." Jensen's law holds: the more heads you train, the more heads you save. A 35B multimodal variant arriving before this one's embeddings settle seems ambitious, but that's the scaling ladder for you. One thing the card doesn't answer: what happens to that 9× latency claim when the candidate set isn't ~1k but, say, 73?