Nous Research: at the time we release Token Superposition Training (TST), a modification to the standard LLM pretraining loop that produces...

at the time we release Token Superposition Training, a modification to the standard LLM pretraining loop at the time we release Token Superposition Training, a modification to the standard LLM pretraining loop that produces a 2-3× wall-clock speedup at matched FLOPs without changing the model architecture, optimizer, tokenizer, or training data. During the first third of training, the model reads and predicts contiguous bags of tokens, averaging their embeddings on the input side and predicting the next bag with a modified cross-entropy on the output side. For the remainder of the run, it trains normally on next-token prediction. The inference-time model is identical to one produced by conventional pretraining. The BotsFired read on Nous Research: this matters only where it changes a real workflow, budget, deployment constraint, or buyer decision. Model progress only matters when it changes the work people can actually ship. The hard question is whether the model improves accuracy, latency, cost, or tool use enough to change a production decision. For builders, the practical question is whether Nous Research removes a constraint that teams already feel. A launch is cheap; a default workflow is expensive to earn. Developers will test it against real tasks, not leaderboard language: messy prompts, long context, tool calls, regressions, and the cost of switching. For operators, the test is narrower: what work moves from manual judgment to repeatable system behavior, who approves the change, and what failure mode appears when the tool leaves the demo path?