Poolside has released Laguna S 2.1, an open-weight coding model with 118 billion total parameters that activates only 8 billion of them for any given token. The company published the model card on Hugging Face on July 21, 2026, positioning the release as the middle entry in a three-model family built for agentic software engineering. Poolside had already shipped the smallest member, Laguna XS 2.1 (33 billion total, 3 billion active), and Laguna S 2.1 is the step up in that same lineup, not a rerun of it; a larger 225 billion parameter model, Laguna M.1, sits above both.
The mixture-of-experts design is the economic story here. A dense model with 118 billion parameters would normally demand the memory and compute budget of a model that size at inference. Laguna S 2.1 instead routes each token through 10 of its 256 specialized experts plus one shared expert, activating roughly 8 billion parameters per forward pass regardless of the model’s total size. That ratio, roughly 14 to 1 between total and active parameters, is why Poolside can serve the model close to an 8 billion parameter dense model’s cost while drawing on a far larger pool of specialized weights.
The architecture pairs that routing scheme with a 1,048,576 token context window and native reasoning support, meaning the model reasons between tool calls and can carry that reasoning across turns using an enable_thinking control. The combination targets a specific failure mode in agentic coding. Long-horizon tasks require an agent to hold an entire codebase, prior tool outputs, and its own reasoning trace across many sequential tool calls without truncating history. A million-token window without reasoning between steps still loses track of multi-step plans; reasoning without long context still forgets what the agent already tried.
Poolside’s model card reports scores on six benchmarks, including Terminal-Bench 2.1, SWE-bench Multilingual, and Toolathlon Verified, and the comparisons cut both ways. On Terminal-Bench 2.1, Laguna S 2.1 scores 70.2 percent, behind Tencent’s 295 billion parameter Hy3 (71.7 percent) and well behind Anthropic’s Claude Fable 5 (88 percent) and Kimi K3, a 2.8 trillion parameter model, at 88.3 percent. On SWE-bench Multilingual, Laguna S 2.1’s 78.5 percent leads every other model in the table that reports a score there, including systems many times its active size. The pattern is a model with an 8 billion parameter inference footprint trading wins and losses with systems ten to twenty times larger, not a model that sweeps the board.
The license is the other structural choice. Poolside released Laguna S 2.1 under OpenMDW-1.1, which its card describes as permitting free use and modification of the model and associated materials for both commercial and non-commercial purposes. That is a broader grant than open-weight licenses that gate commercial use behind a separate agreement or cap free use above a revenue or usage threshold. Poolside has not published independent benchmark verification beyond the third-party sources it cites for a subset of scores (Artificial Analysis, Scale AI, and Toolathlon’s own leaderboard); the rest of the comparison table is Poolside’s self-reported testing, not a neutral evaluator running all models under identical conditions.
For teams already running Laguna XS 2.1 in production coding agents, Laguna S 2.1 is the upgrade to test first. It shares the same tokenizer and the same engine support across vLLM, SGLang, and llama.cpp, and Poolside shipped a dedicated DFlash speculative decoding model built to keep the larger model’s latency close to its smaller sibling’s, which is the real question to benchmark before committing a production agent to the bigger checkpoint.
Released by Poolside on July 21, 2026, per the model card Poolside published on Hugging Face.
