--- title: Poolside releases Laguna S 2.1, a 118B sparse MoE model... description: Poolside has released Laguna S 2.1, an open-weight sparse mixture-of-experts model with 118B total parameters (8B active per token) designed for agentic coding... canonical: https://daily.dev/posts/poolside-releases-laguna-s-2-1-a-118b-sparse-moe-model-for-agentic-coding-0xmswtm3e twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Poolside releases Laguna S 2.1, a 118B sparse MoE model for agentic coding | daily.dev og:description: Poolside has released Laguna S 2.1, an open-weight sparse mixture-of-experts model with 118B total parameters (8B active per token) designed for agentic coding... og:url: https://daily.dev/posts/poolside-releases-laguna-s-2-1-a-118b-sparse-moe-model-for-agentic-coding-0xmswtm3e og:image: https://api.daily.dev/og/posts/0XmswTm3e.png og:image:alt: Poolside releases Laguna S 2.1, a 118B sparse MoE model for agentic coding og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Poolside releases Laguna S 2.1, a 118B sparse MoE model for agentic coding **[Collections](https://daily.dev/sources/collections)** · 3 min read · 1 upvotes · 0 comments ## Summary Poolside has released Laguna S 2.1, an open-weight sparse mixture-of-experts model with 118B total parameters (8B active per token) designed for agentic coding and long-horizon tasks. It features a 1M token context window, supports thinking and no-thinking modes, and is licensed under OpenMDW-1.1. The model is optimized for multi-step agentic runs including planning, tool calling, error recovery, and self-checking. It runs on a single NVIDIA DGX Spark via NVFP4 quantization, integrates with vLLM out of the box, and is backward-compatible with existing Laguna serve setups using the poolside_v1 parsers. ## Content ## What it is Poolside has released Laguna S 2.1, a 118B-parameter sparse Mixture-of-Experts model that activates only 8B parameters per token. It supports a 1M-token context window, runs in both thinking and no-thinking modes, and is available on Hugging Face under the Linux Foundation's OpenMDW-1.1 license. The model was built in under nine weeks on 4,000 H200 GPUs using Poolside's internal Model Factory platform, which takes a model from pre-training to release in five to eight weeks. ## Benchmark performance Despite its relatively small active parameter count, Laguna S 2.1 holds up well against much larger models: - **Terminal-Bench 2.1:** 70.2% - **SWE-Bench Pro:** ~60% - **SWE-Bench Multilingual:** 78.5% In a side-by-side game coding test on atomic.chat, it matched the 753B GLM-5.2 while using 10.3K tokens versus GLM-5.2's 26.4K. It still trails closed-source leaders from OpenAI and Anthropic by roughly 10–15 percentage points. ## What makes it different Poolside's co-CEO Eiso Kant is explicit that the gains here aren't from raw scale. The improvements come from behavioral qualities: persistence across long task sequences, self-verification, and the ability to backtrack and recover from failures. Training changes include more generous rollout budgets during RL, multi-harness rollouts to reduce scaffold overfitting, and FP8 RL training. The model can autonomously work on tasks for up to 24 hours, using tools, checking its own work, and continuing without human intervention. Documented examples include building a browser rendering engine over 181 reasoning turns, optimizing a Go codebase for a 5.2% speedup and 70% memory reduction, and independently re-deriving a proof to Erdős problem #397. ## Running it locally The official NVFP4 quantization runs on a single Nvidia DGX Spark desktop. For Mac users, mixed q2/q3 quants are available that run on MacBook systems with 64GB RAM, with generation speeds around 65 tokens/sec and prefill around 570 tokens/sec using DwarfStar or llama.cpp. A Metal-only implementation in the DwarfStar laguna-s2.1 branch hits roughly 50 tokens/sec generation and 500 tokens/sec prefill on an M5 Max. vLLM runs it out of the box. Existing Laguna serve setups carry over, and the `poolside_v1` tool-call and reasoning parsers already work. It's also available on OpenRouter, Vercel AI Gateway, Kilo (free for a limited time), and OpenCode. ## The Western open-weight angle Poolside, backed by Nvidia and eBay at a $3B valuation, frames this release as a Western alternative to DeepSeek and Qwen for enterprises and governments that want to self-host without routing data to foreign providers. The weights are open, the eval dataset is fully published with six public benchmarks and four runs each, and the company runs roughly 10,000–20,000 experiments per month with a team that's 90% engineering and 10% research. Poolside has also been unusually transparent about methodology, publishing papers and exposing full eval runs so users can check for reward hacking themselves. ## Questions this post answers ### What are the benchmark scores for Poolside's Laguna S 2.1 model on agentic coding tasks? Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, approximately 60% on SWE-Bench Pro, and 78.5% on SWE-Bench Multilingual. It's a 118B-parameter sparse Mixture-of-Experts model activating only 8B parameters per token, and while it still trails closed-source leaders from OpenAI and Anthropic by roughly 10-15 percentage points, it matched the much larger 753B GLM-5.2 in a side-by-side game coding test while using far fewer tokens (10.3K versus 26.4K). _Comparing agentic coding models on real benchmarks gets easier when daily.dev surfaces releases like Laguna S 2.1 alongside the rest._ ### Can I run Poolside's Laguna S 2.1 model locally on a Mac? Yes, mixed q2/q3 quantizations of Laguna S 2.1 run on MacBook systems with 64GB RAM, achieving around 65 tokens/sec generation and 570 tokens/sec prefill using DwarfStar or llama.cpp. A Metal-only implementation in the DwarfStar laguna-s2.1 branch reaches roughly 50 tokens/sec generation and 500 tokens/sec prefill on an M5 Max. The official NVFP4 quantization also runs on a single Nvidia DGX Spark desktop. _Developers weighing local versus cloud inference for coding models can track setups like this one on daily.dev._ ### What license does Poolside use for the Laguna S 2.1 model weights? Laguna S 2.1 is released under the Linux Foundation's OpenMDW-1.1 license and is available on Hugging Face. Poolside frames the open-weight release as a Western alternative to DeepSeek and Qwen for enterprises and governments wanting to self-host models without routing data to foreign providers, and has published the full eval dataset with six public benchmarks and four runs each. _Teams weighing open-weight licensing options for self-hosted coding models can follow releases like this on daily.dev._ --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-coding](https://daily.dev/tags/ai-coding), [#vllm](https://daily.dev/tags/vllm), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/poolside-releases-laguna-s-2-1-a-118b-sparse-moe-model-for-agentic-coding-0xmswtm3e) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Poolside releases Laguna S 2.1, a 118B sparse MoE model for agentic coding","url":"https://daily.dev/posts/poolside-releases-laguna-s-2-1-a-118b-sparse-moe-model-for-agentic-coding-0xmswtm3e","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/poolside-releases-laguna-s-2-1-a-118b-sparse-moe-model-for-agentic-coding-0xmswtm3e"},"datePublished":"2026-07-21T17:19:22.455Z","dateModified":"2026-09-13T19:38:38.001Z","description":"Poolside has released Laguna S 2.1, an open-weight sparse mixture-of-experts model with 118B total parameters (8B active per token) designed for agentic coding...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/poolside-releases-laguna-s-2-1-a-118b-sparse-moe-model-for-agentic-coding-0xmswtm3e","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-coding,vllm,mixture-of-experts","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Poolside releases Laguna S 2.1, a 118B sparse MoE model for agentic coding"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/poolside-releases-laguna-s-2-1-a-118b-sparse-moe-model-for-agentic-coding-0xmswtm3e#faq","mainEntity":[{"@type":"Question","name":"What are the benchmark scores for Poolside's Laguna S 2.1 model on agentic coding tasks?","acceptedAnswer":{"@type":"Answer","text":"Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, approximately 60% on SWE-Bench Pro, and 78.5% on SWE-Bench Multilingual. It's a 118B-parameter sparse Mixture-of-Experts model activating only 8B parameters per token, and while it still trails closed-source leaders from OpenAI and Anthropic by roughly 10-15 percentage points, it matched the much larger 753B GLM-5.2 in a side-by-side game coding test while using far fewer tokens (10.3K versus 26.4K). Comparing agentic coding models on real benchmarks gets easier when daily.dev surfaces releases like Laguna S 2.1 alongside the rest."}},{"@type":"Question","name":"Can I run Poolside's Laguna S 2.1 model locally on a Mac?","acceptedAnswer":{"@type":"Answer","text":"Yes, mixed q2/q3 quantizations of Laguna S 2.1 run on MacBook systems with 64GB RAM, achieving around 65 tokens/sec generation and 570 tokens/sec prefill using DwarfStar or llama.cpp. A Metal-only implementation in the DwarfStar laguna-s2.1 branch reaches roughly 50 tokens/sec generation and 500 tokens/sec prefill on an M5 Max. The official NVFP4 quantization also runs on a single Nvidia DGX Spark desktop. Developers weighing local versus cloud inference for coding models can track setups like this one on daily.dev."}},{"@type":"Question","name":"What license does Poolside use for the Laguna S 2.1 model weights?","acceptedAnswer":{"@type":"Answer","text":"Laguna S 2.1 is released under the Linux Foundation's OpenMDW-1.1 license and is available on Hugging Face. Poolside frames the open-weight release as a Western alternative to DeepSeek and Qwen for enterprises and governments wanting to self-host models without routing data to foreign providers, and has published the full eval dataset with six public benchmarks and four runs each. Teams weighing open-weight licensing options for self-hosted coding models can follow releases like this on daily.dev."}}]} ```