Laguna S 2.1 posts numbers that sit next to models five to twenty times its size, but the leaderboard row isn’t the part worth your attention. The Poolside writeup reports a 118B-parameter Mixture-of-Experts model that activates only 8B parameters per token, went from cold start to release in under nine weeks, and carries a 1M-token context in both thinking and no-thinking modes.

If you run agentic coding loops in production, active-param count is the line item that actually shows up on your latency budget and your bill — not total parameter count, and not the SWE-Bench headline. An 8B-active model that resolves 70.2 on Terminal-Bench 2.1 and 78.5 on SWE-Bench Multilingual is buying you frontier-adjacent task completion at a serving cost a 1.6T-param model can’t touch. That trade — small active footprint, long horizon, competitive resolution rate — is the one I care about when I’m sizing an orchestration tier that has to fan out dozens of tool calls per task.

The move I respect most is releasing full trajectories for every trial in the final eval set at trajectories.poolside.ai. Benchmark scores are a summary statistic; trajectories are what an eval-minded team can audit for reward hacking, tool-call thrash, and the failure modes a pass@1 number hides. Most vendors publish the score and bury the runs.

The under-nine-weeks cadence is the quieter signal. Three models in three months means the moat isn’t any single checkpoint — it’s the training loop that keeps producing them. The HN thread is arguing about benchmark honesty, which misses it.

So here’s the question I’d put to anyone shipping agents: when you last swapped models, did you compare active-parameter serving cost, or did you compare leaderboard rows?