Laguna Coding Models by Poolside: status and timeline

Milestones

Next · Enterprise deployments and government contracts

Laguna XS.2 and Laguna M.1 released28 Apr 2026Complete.
Laguna XS 2.1 released2 Jul 2026Complete.
Laguna S 2.1 released and tops SWE-Bench Multilingual21 Jul 2026Complete.
Series C funding at $12 billion pre-money valuation announcedAug 2026Complete.
Enterprise deployments and government contractsTarget 2027Not yet reached.

Most important updates

Upcoming

2027Enterprise deployments and government contracts (next)

Current obstacles

Scaling to match closed models on complex reasoningLarger closed models still outperform Laguna on tasks requiring multi-step physical or logical reasoning beyond code.
Governance and safety in code generationDefense and enterprise customers require guarantees about generated code safety, which open models struggle to provide.

Physics limits

Transformer scaling limitsBeyond ~200B parameters, dense transformers hit efficiency and training time walls; sparse routing helps but eventual dense-to-sparse transitions remain complex.
Training data quality for codingCode quality in training data directly affects output; proprietary codebases and real-world bug patterns remain hard to acquire and license at scale.

How it works

3 parts

Sparse routing

256 routed experts with selective activation

Only 8B of 118B parameters activate per token, reducing computation while maintaining capability. Uses grouped-query attention with 8 KV heads.

Long context

Full 1M token context always available

Unlike some models that truncate context in reasoning mode, Laguna S 2.1 maintains the full context for both thinking and direct modes.

FP8 training

Efficiency through low-precision RL

Reinforcement learning entirely in FP8 precision reduces memory and training cost, a first for this scale of model.

Spec sheet

Spec	Laguna Coding Models
Laguna S 2.1 total parameters	118 billion^R (reported)
Laguna S 2.1 active parameters per token	8 billion^R (reported)
SWE-Bench Multilingual score	78.5%^R (reported)
Context window	1,048,576 tokens^R (reported)
Training compute	4,000 NVIDIA H200 GPUs in under 4 weeks^R (reported)
Laguna M.1 total parameters	225 billion^R (reported)

^R reported by the company

Papers & demos

1 item

21 Jul 2026demo
Laguna S 2.1 release (opens rits.shanghai.nyu.edu)Open-weight 118B coding model; 78.5% on SWE-Bench Multilingual; demonstrates frontier performance in open coding models.

Update log

6 updates

About Poolside

The team behind Laguna Coding Models

Poolside

Coding agents, Laguna models, agentic coding

Poolside builds large language models specialized for agentic software coding, deployed in enterprise and government environments. Founded by former GitHub CTO Jason Warner, the company released Laguna S 2.1, a 118-billion-parameter open-weight model that tops coding benchmarks.

Series C August 2026: Nvidia invests $1B and licenses Model Factory for $6B at $12B valuation.
Laguna S 2.1, released July 21, 2026, achieves 78.5% on SWE-Bench Multilingual, outperforming models 10–20x its size.
The model activates 8 billion of 118 billion parameters per token and fits on a single NVIDIA DGX Spark when quantized.

poolside.ai (opens poolside.ai)Company profile →

Founded: 20233 yrs
Headquarters: United States
Status: Private
Valuation: $12BprivateAug 2026
Raised: $1.6B4 rounds
Last round: Series C · $1BAug 2026
Works in: AICoding agents
Coverage: 1 program · 6 updateslatest 1 Aug 2026checked 11 Oct
Lead investors: NvidiaBain Capital VenturesDST GlobalFelicis VenturesXavier NielRedpoint Ventures
Partners: Amazon Web Services (opens aws.amazon.com)
Investors: Nvidia (opens nvidia.com)
People: Jason Warner (opens en.wikipedia.org)CEO & Co-founderEiso Kant (opens linkedin.com)Co-founder

Spec

Laguna Coding Models

Laguna S 2.1 total parameters

118 billion^R (reported)

Laguna S 2.1 active parameters per token

8 billion^R (reported)

SWE-Bench Multilingual score

78.5%^R (reported)

Context window

1,048,576 tokens^R (reported)

Training compute

4,000 NVIDIA H200 GPUs in under 4 weeks^R (reported)

Laguna M.1 total parameters

225 billion^R (reported)

Laguna Coding Models

Milestones

Most important updates

Upcoming

Current obstacles

Physics limits

How it works

256 routed experts with selective activation

Full 1M token context always available

Efficiency through low-precision RL

Spec sheet

Papers & demos

Update log

Aug 2026

Tue 21 Jul

Thu 2 Jul

Tue 28 Apr

Dec 2024

Oct 2024

About Poolside

Current obstacles

Physics limits

How it works

256 routed experts with selective activation

Full 1M token context always available

Efficiency through low-precision RL

Spec sheet

Papers & demos

Update log

Aug 2026

Tue 21 Jul

Thu 2 Jul

Tue 28 Apr

Dec 2024

Oct 2024

About Poolside