Poolside

Overview

AI lab building models for agentic coding. Released Laguna S 2.1 on 2026-07-21 as an open-weight model under the OpenMDW-1.1 license (source).

The release's argument is about size rather than scale: a 118B total / 8B active MoE reported to beat models several times larger on long-horizon coding benchmarks (source).

Key People

unknown — not obtainable from this environment when this page was created.

Models & Products

• Laguna S 2.1 — released 2026-07-21. 118B total / 8B active MoE, up to 1M context in thinking and no-thinking modes, OpenMDW-1.1 (source)

Recent Activity

• 2026-07-21: Laguna S 2.1 released — open-weight agentic coding model, 118B total / 8B active MoE, up to 1M context in both thinking and no-thinking modes, weights on Hugging Face under OpenMDW-1.1 (use, modification and redistribution permitted, including commercially). Poolside's published figures: 70.2% on Terminal-Bench 2.1 with thinking, 78.5% on SWE-Bench Multilingual, 40.4% on DeepSWE v1.1 against DeepSeek-V4-Pro-Max's 9.0% at roughly one-sixth the active parameters. → Laguna S 2.1 (source)

• 2026-08-02: Named as one of three open models on the Pareto frontier — Interconnects' "Latest open artifacts (#23)" names Laguna S2.1 alongside Inkling and Kimi K3 (source).

• 2026-08-20: NVIDIA licenses Poolside's "Model Factory", hires 109 staff, invests $1B — founders stay. NVIDIA is reported to pay ~$6 billion to license the "Model Factory" (Poolside's data-processing, training, RL and evaluation suite) and to hire 109 Poolside staff, plus a $1 billion investment at a $12 billion pre-money valuation. Described as "not an acquisition and not an acquihire" — a reverse acquihire in which the three founders remain and the company continues; coverage notes NVIDIA has used the same structure twice before. Poolside's infrastructure spin-out, the Poolside Infrastructure Company (PIC) (spun out Jan 2026), is building a 1.2GW Texas datacenter described as scaling to a 7GW neocloud. Held at reporting confidence: assembled from secondary coverage (Bloomberg/Newcomer, The Next Web, The Decoder), not a first-party statement, and the value framing varies by outlet (source).

Strategic Position

• The framing in coverage is explicitly geographic: The Next Web describes Laguna S 2.1 as "the open-weight coding model pitched as the West's answer to DeepSeek and Qwen" (source)
• Competes on active parameters per unit of capability rather than total size — the DeepSWE v1.1 comparison Poolside chose to publish is against a model with roughly six times the active parameters (source)
• OpenMDW-1.1 is not one of the licenses this wiki has seen before on a frontier release; the field's Western open-weight releases have otherwise been Apache 2.0 (Inkling, Mistral Large 3) and the Chinese ones MIT (DeepSeek V4-Flash)
• Poolside does not appear on either leaderboard snapshot this repo holds (source), so every figure on Laguna S 2.1 is currently vendor-run
• The NVIDIA deal reframes the lab from a model vendor to a compute play. The transaction's value sits in the Model Factory infrastructure and a 7GW neocloud, not in Laguna S2.1 — NVIDIA licenses the machinery for building models and takes the researchers, while the model business continues under the founders. It is the second Western open-weight coding lab this window to be absorbed into a larger compute story rather than compete as a standalone product (source)

Related

• Laguna S 2.1
• NVIDIA — licensed Poolside's Model Factory and hired 109 staff (2026-08-20)
• Thinking Machines Lab — the other Western open-weight release this window
• Open-Weights Policy Fight
• Agentic Reinforcement Learning