Laguna XS 2.1

Poolside officially introduced Laguna XS 2.1 on July 2, 2026, as an upgraded version of its Laguna XS.2 coding model. Per the company's blog post, Laguna XS 2.1 is a 33B total-parameter Mixture-of-Experts model with 3B activated parameters per token, designed for agentic coding and long-horizon work on a local machine, The release emphasizes local deployment: Laguna XS 2.1 ships with launch-day support in vLLM, SGLang, NVIDIA TensorRT-LLM, Hugging Face transformers, and Ollama, with llama.cpp support coming soon. Poolside is publishing FP8, INT4, and NVFP4 quantized checkpoints, plans GGUF checkpoints alongside llama.cpp, and is open

The GPU Trade reported on July 5, 2026 that Poolside launched Laguna XS 2.1 on July 2, 2026 as an open-weight 33B MoE aimed at agentic coding and local inference. The piece cites the 33B total / ~3B activated parameters, a 262,144-token context window, 40 layers, and 256 experts, and highlights a 5.4-point jump on SWE- On-device efficiency is a focus of the coverage: Poolside and the Hugging Face listing point to KV-cache quantization to FP8, launch-day support in vLLM and TRT-LLM, and multiple quantized checkpoints (FP8, INT4, NVFP4) to reduce VRAM and accelerate inference. The story frames Laguna XS 2.1 as part of Poolside's broade

AIDeveloper44 published a technical deep-dive on July 4, 2026 covering Poolside's Laguna XS 2.1 release for local agentic coding. The piece describes the 33B total / 3B active MoE design with 256 individual experts plus one shared expert, routing each token through the top eight most relevant paths, and confirms the 25 On deployment, the article lists FP8, INT4, NVFP4, and GGUF checkpoints for local execution and highlights an open-weight 0.5B DFlash draft model for speculative decoding, citing measured speedups of 1.67x to 2.64x. The licensing is described as the Linux Foundation's permissive OpenMDW-1.1 framework, and the overall p

llm-stats lists Laguna XS 2.1 at an overall rank of 169 in its composite LLM Stats Score, with capability tiers of "Average / Top half" for reasoning (163 of 369) and "Bad / Below top half" for coding (138 of 273) and tool calling (187 of 200). The page shows a blended price around $0.10 per million tokens, positioning The benchmark table attributes scores to Poolside's own evaluation harness: SWE-bench Verified at 0.71 (rank 62), SWE-bench Multilingual at 0.63 (rank 33), and SWE-Bench Pro at 0.48 (rank 56), all using "Harbor + Poolside agent harness; thinking on; 256K ctx; temp=1.0, top_k=20, top_p=1" with mean pass@1 over 4 attempt

The Kilo Code model profile for poolside/laguna-xs-2.1, dated July 2, 2026, lists the model as the latest coding agent in the 33B-A3B category and a step forward from the April 2026 Laguna XS.2 release. It reports a 262,144-token context window, a 32,768-token maximum output, text input only, and pricing via OpenRouter For evaluation evidence, Kilo Code cites a Kilo Bench score of 26.7% completion on Terminal Bench 2.0 at an average cost of $12.03 per complete benchmark attempt, labeled "Official Kilo eval results," and shows weekly token usage charts with no leaderboard ranking data yet. Because the page is a third-party aggregator

Poolside's official release notes explicitly list Laguna XS 2.1 as a model release, describing it as an incremental update to Laguna XS.2 that adds native reasoning support and improves performance on multilingual coding and terminal-style tasks. The notes confirm the model supports a 256K context window and is positio The release notes index situates Laguna XS 2.1 within Poolside's broader model lineup, alongside the April 2026 releases of Laguna M.1 and Laguna XS.2, and the earlier Malibu 2.x series. This first-party versioning record confirms creator attribution and the model's lineage as part of Poolside's agentic-coding family,

The Explainx Substack newsletter from July 5, 2026 reports that Poolside launched Laguna XS 2.1 as an upgraded open coding model with a 262K-token context window, built on a 33B-parameter Mixture-of-Experts architecture that activates about 3B parameters per token. It frames the release as aimed at agentic software dev As a secondary newsletter, the article paraphrases vendor materials rather than providing original benchmarks, citations, or independent testing, so its architecture and deployment claims should be read as corroborating context rather than primary evidence. It also does not reconcile the "local deployment" framing with