Poolside's Laguna S 2.1: A Small Coding Model That Publishes Its Own Trajectories
Laguna S 2.1 is a 118B-parameter model activating 8B parameters per token, scoring 70.2% on Terminal-Bench 2.1 with every trajectory public.
What Laguna S 2.1 actually is
Poolside describes Laguna S 2.1 as a 118B total parameter Mixture-of-Experts model with 8B activated parameters per token, supporting a context window of up to 1M tokens in both thinking and no-thinking modes. According to Poolside's announcement, it went from the start of training to launch in under nine weeks, and the company positions it as, in its own words, "as far as we can measure, the most capable agentic coding model in its weight class by a wide margin."
The benchmark scores, and why Poolside publishes trajectories
On Terminal-Bench 2.1, Laguna S 2.1 scores 70.2% with thinking enabled in Poolside's own agent harness. On the harder DeepSWE v1.1 benchmark, where frontier models range from 54% to 73% and some models with far more parameters score below 10%, Laguna S 2.1 scores 40.4 in thinking mode. Poolside says it computes every benchmark as the maximum of the vendor's own self-reported score, a benchmark author's leaderboard, or a third-party leaderboard, and for every score it publishes, it also publishes the full trajectory of every trial in the final evaluation set at trajectories.poolside.ai, so outsiders can inspect exactly how the model reached its answer rather than trusting a number alone.
What the model did in real tasks
Poolside shares three worked examples with unedited trajectories. In one, asked to build a browser engine from scratch in JavaScript, Laguna S 2.1 worked for one 50-minute session of 181 steps with no human help, building a working HTML/CSS rendering engine and then checking its own work by rendering the same markup in a real browser and comparing screenshots. In another, a researcher pointed the model at Poolside's own training-and-evaluation harness and asked it to optimize speed and memory. Over multiple hours, the model found and applied several targeted fixes, replacing costly string concatenation with buffers and cutting redundant copying, and ended up making the harness 5.2% faster while reducing memory allocation by about 70%.
Where scores get harder to trust
Poolside is candid about a broader problem with the field: as benchmarks mature, top scores cluster tightly, and models that behave very differently end up only a few points apart. The company separately flags that evaluating agent models is notoriously difficult because of reward hacking, and says recent work has focused on adversarial judging to catch more of it.
Key facts
- Laguna S 2.1: 118B total parameters, 8B activated per token, up to 1M tokens of context, released by Poolside.
- Scores 70.2% on Terminal-Bench 2.1 and 40.4 on DeepSWE v1.1, both with thinking enabled.
- Poolside publishes full evaluation trajectories for every trial at trajectories.poolside.ai.
- In one real-task demo, the model built a browser engine in one 50-minute, 181-step session with no human intervention.
- Optimizing its own training harness, the model achieved a 5.2% speedup and cut memory allocation by about 70%.
Check your understanding
- Why does Poolside publish full trajectories alongside its benchmark scores instead of just the scores themselves?
- How does Laguna S 2.1's 40.4 score on DeepSWE v1.1 compare with the 54% to 73% range Poolside reports for frontier models on the same benchmark?
- Why does Poolside say evaluating agent models is notoriously difficult, and what approach does it say it uses to catch reward hacking?