HybridLink on Two Pascal GPUs: Tokens per Second Was the Wrong Metric Running Together's HybridAgent checkpoints on two 2016 TITAN X Pascal GPUs, a same-checkpoint on/off comparison with an external grader found that enabling HybridLink gave the Qwen3-8B-Base model a median 2.01× speedup across the eight pairs where both modes finished correctly and raised correct completions from 11/18 to 13/18, while the 14B model's median paired speedup was 1.00× but its correct completions rose from 7/18 to 12/18. Time to first reasoning text did not move at either size and visible tokens per second fell in both cases, which the author argues makes tokens per second the wrong metric for latent reasoning. The author also found that community GGUF conversions through Ollama never load the HybridLink adapter, so their rates (about 53.0 tok/s for the 8B and 23.6 tok/s for the 14B at Q4_K_M) measure only the backbone. HybridLink on Two Pascal GPUs: Tokens per Second Was the Wrong Metric $ grep -n "^ " 2026-10-hybridlink-two-pascal-gpus-tokens-per-second.md I spent a day running Together's HybridAgent https://arxiv.org/abs/2610.18861v1 checkpoints on two 2016 TITAN X Pascal cards, trying to find out whether latent reasoning gets a local model to a correct answer sooner. My first metric was tokens per second. It could not answer the question, and my first set of numbers came from a model with the interesting part switched off. The measurement that could answer it was a same-checkpoint on/off comparison with an external grader. On six fixed arithmetic questions, three seeds each, enabling HybridLink gave the 8B model a median 2.01× speedup across the eight pairs where both modes finished correctly, and raised correct completions from 11/18 to 13/18. The 14B model's median paired speedup was 1.00×, but it went from 7/18 correct to 12/18. Time to first reasoning text did not move at either size. Visible tokens per second went down in both cases. Those three numbers answer different questions, and conflating them is how a latent-reasoning result gets misreported in either direction. What HybridLink does in the released code what-hybridlink-does-in-the-released-code The paper's idea is that a reasoning model should not have to put every intermediate thought through the vocabulary. Inside a