Poolside.ai · Laguna Model Family · Research Intelligence Report
The West's Most
Capable Open Coding AI
50+ verified data points on Poolside's Laguna model family — architecture, benchmarks, training infrastructure, pricing, and competitive positioning as of July 2026.
As of July 22, 2026 · Data verified from primary sources
Model Architecture Metrics
Laguna S 2.1 · Total Parameters
118B
Total parameters with only 8B active per token via MoE routing
Laguna M.1 · Total Parameters
225B
23B activated parameters — trained on 6,144 NVIDIA H200 GPUs
Laguna XS.2 · Total Parameters
33B
Only 3B activated per token; runs on a Mac with 36 GB RAM via Ollama
Laguna S 2.1 · Max Context
1M
1,048,576-token context window; default 256K on OpenRouter
Training Data Scale
30T+
Both M.1 and XS.2 pre-trained on 30+ trillion tokens from scratch
S 2.1 Training Speed
<9wk
Under 9 weeks from training start to public release; <4 weeks on 4,000 H200s
Benchmark Performance Comparison
| Model | SWE-Bench Verified | SWE-Bench Multilingual | SWE-Bench Pro | Terminal-Bench |
|---|---|---|---|---|
| Laguna S 2.1 Laguna S 2.1 Jul 2026 · 118B/8B MoE | — |
78.5% |
59.4% |
70.2% |
| Laguna M.1 Laguna M.1 Apr 2026 · 225B/23B MoE | 72.5% |
57.4% |
46.9% |
40.7% |
| Laguna XS 2.1 Laguna XS 2.1 Jul 2, 2026 · 33B/3B MoE | 70.9% |
63.1% |
— |
37.5% |
| Laguna XS.2 Laguna XS.2 Apr 2026 · 33B/3B MoE · Apache 2.0 | 68.2% |
62.4% |
44.5% |
30.1% |
★ Terminal-Bench version: 2.0 for M.1/XS.2/XS 2.1; Terminal-Bench 2.1 for Laguna S 2.1. SWE-Bench Verified baseline (prior Malibu 2.2 model): 55.6%. Sources: poolside.ai/blog, docs.poolside.ai
Company & Infrastructure Numbers
$626M
Total Funding Raised
$3B
Series B Valuation (Oct 2024)
$12B
Target Series C Valuation (2025)
2GW
Project Horizon AI Campus (Texas)
40K+
GB300 NVL72 GPUs (CoreWeave deal)
60
Applied Research Team Members
~$50M
Estimated Annual Revenue
2023
Founded (San Francisco)
Training Efficiency & Infrastructure
15%
Muon Optimizer Edge over AdamW — Poolside's distributed Muon implementation achieves the same training loss in ~15% fewer steps than the standard AdamW optimizer, while requiring only one state per parameter (vs. two for AdamW), cutting checkpoint memory.
6,144
GPUs for Laguna M.1 Pre-training — Laguna M.1 was trained using 6,144 interconnected NVIDIA Hopper (H200) GPUs with NVLink; Laguna XS.2 used 2,048 H200 GPUs.
4,000
H200 GPUs for Laguna S 2.1 — Pre-training for the S 2.1 model began May 22, 2026 on 4,000 NVIDIA H200 GPUs, completing the entire cycle in under 4 weeks.
<1%
Muon Optimizer Overhead — During Laguna M.1 pre-training, the combined Muon optimizer overhead (including Newton–Schulz iterations) was reduced to under 1% of total training step time.
13%
Synthetic Data in Training Mix — Approximately 13% of Laguna XS.2's final training mix consists of synthetic data, with the broader Laguna family having used 4.4T+ synthetic tokens in total.
~60
AutoMixer Proxy Models — Instead of hand-tuned data mixing ratios, Poolside trained ~60 proxy models on different data compositions to automatically optimize training data mix via learned regression.
6.8%
Laguna S 2.1 Parameter Activation Rate — Laguna S 2.1 activates only ~6.8% of its 118B parameters on any given token, enabling frontier-class performance on hardware as compact as a single DGX Spark.
Laguna Family Release Timeline
October 2024
Series B · $500M raised at $3B valuation
Bain Capital Ventures leads round with participation from Nvidia, eBay Ventures, DST Global. Company becomes unicorn within 18 months of founding.
December 2024
AWS Strategic Partnership Announced
Poolside becomes first-party AWS offering. Models deployable on Amazon Bedrock and EC2 within customer security boundaries.
November 2025
Fern Labs Acquisition
Poolside acquires London-based Fern Labs, creators of Bridge — a multi-agent orchestration layer for high-stakes production environments.
April 28, 2026
Laguna M.1 + XS.2 Public Launch
Two foundation models released: 225B M.1 (closed) and 33B XS.2 (Apache 2.0 open-weight). Both appear on Hugging Face and OpenRouter same day. SWE-bench Verified: 72.5% (M.1) and 68.2% (XS.2).
July 2, 2026
Laguna XS 2.1 Released
Incremental update to XS.2 adds native reasoning support. SWE-bench Multilingual improves +5.4% to 63.1%. License upgraded to OpenMDW-1.1. Old XS.2 API retired July 9, 2026.
July 21, 2026
Laguna S 2.1 Released — West's Most Capable Open Model
118B/8B MoE model tops Terminal-Bench 2.1 at 70.2%, SWE-Bench Multilingual at 78.5%, DeepSWE at 40.4%. Runs on single DGX Spark. Weights on Hugging Face under OpenMDW-1.1.
API Pricing (OpenRouter, July 2026)
Laguna XS 2.1
$0.06
per 1M input tokens
Output: $0.12/M
Context: 256K · Free tier available
Laguna M.1
$0.20
per 1M input tokens
Output: $0.40/M
Context: 256K · 32K max output
Laguna S 2.1
$0.10
per 1M input tokens (256K)
Output: $0.20/M
1M context tier also available · Free tier
Prompt caching on OpenRouter can reduce effective costs by 60–80% depending on repeated context volume.