Most large models either optimize for pure inference throughput or for dense-capacity reasoning; Laguna XS.2 deliberately trades dense parameter activation for a sparse Mixture-of-Experts layout to make multi-step, tool-enabled coding and very long-context reasoning practical on a single workstation.
Key Capabilities
- Sparse MoE architecture: 33B total parameters with ~3B activated per token and 256 experts, which lowers runtime memory while keeping high capacity for diverse coding tasks.
- Mixed Sliding-Window & Global Attention: 30/10 layer split (3:1) with per-head sigmoid gating and per-layer rotary scales — enables a 131,072-token context window with much lower KV cache costs than fully global attention.
- KV cache quantized to FP8: reduces memory-per-token for long histories, making local inference feasible on machines with ~36 GB RAM (Poolside's guidance).
- Native reasoning / preserved thinking: built-in support for interleaved “thinking” (reasoning) that can be enabled or disabled per request, which aids agentic workflows and tool orchestration.
- Ecosystem support: launch-day integrations with vLLM, Transformers (source install until release), TRT-LLM, and Ollama; offered under Apache-2.0 for commercial use.
Who it's for and tradeoffs
Great fit if you need an LLM that can run locally for sustained, multi-step coding/agent workflows and maintain very long context without renting large GPU clusters. It is particularly useful for terminal-based coding agents, tool-using assistants, and local research experimentation where preserving intermediate reasoning traces matters.
Look elsewhere if you need top leaderboard scores on every code/bench metric (some larger dense models score higher on certain benchmarks), if you require minimal integration complexity (some runtimes require source builds or PR branches), or if you must conform to a non-permissive license.
Where it fits
Laguna XS.2 sits between compact dense open models (which are easy to run but hit capacity limits) and very large multi-hundred-billion models (which require heavy infra). Its MoE design aims to give model-capacity advantages while keeping per-token activation and KV memory low enough for single-machine experiments and local agents.
How it was evaluated (brief)
Poolside reports benchmark results across SWE-bench and Terminal-Bench variants; the model demonstrates competitive pass@1 numbers on several agent benchmarks but shows tradeoffs compared with some 31–35B dense models on selected metrics. Practical deployment notes include guidance for vLLM/Transformers integration and optional TRT-LLM/Ollama workflows for local usage.
