What Is Poolside? The $3B Startup Building the West's Open-Weight Answer to DeepSeek
What Is Poolside? The $3B Startup Building the West’s Open-Weight Answer to DeepSeek
While everyone was watching OpenAI, Anthropic, and Google compete on closed models, a $3 billion startup called Poolside was quietly building the most capable open-weight coding model from a Western company. Their flagship product, Laguna S 2.1, scores 70.2% on Terminal-Bench and 59.4% on SWE-bench Pro while activating only 8B parameters per token.
That combination of open weights, Western origin, and frontier-level coding performance did not exist before Poolside. Here is who they are, how they train differently, and why it matters.
The Company
Poolside is an AI company valued at $3 billion. They focus exclusively on code intelligence. Not general-purpose AI. Not chatbots. Not image generation. Code.
This focus matters because coding models benefit enormously from specialized training infrastructure. While general-purpose labs split their resources across text, vision, audio, and reasoning, Poolside concentrates everything on making models that write better software.
Their hardware allocation reflects this focus: 4,096 H200 GPUs dedicated to training code models. They trained Laguna S 2.1 in under 9 weeks on this cluster, which enables rapid iteration between model versions.
The Product Line
Laguna S 2.1 (Flagship)
Released July 21, 2026. The model that put Poolside on the map for most developers.
- 118B MoE, 8B active per token
- 1M context window (thinking and non-thinking modes)
- 70.2% Terminal-Bench, 59.4% SWE-bench Pro, 40.4% DeepSWE
- Free on OpenRouter at 256K context
- Open weights under OpenMDW-1.1
Full details in our complete Laguna S 2.1 guide.
Laguna XS 2.1 (Consumer)
Released July 2, 2026. The smaller model designed for local deployment.
- 33B MoE, 3B active per token
- 262K context window
- 33.4% Terminal-Bench, 63.1% SWE-bench Multilingual
- Free on OpenRouter
- Runs on a single 24GB GPU at 4-bit quantization
Pool CLI (Agent Tool)
Poolside’s terminal-based coding agent, similar to Claude Code. Connects directly to Laguna models for first-party agentic coding. See our pool CLI guide.
How They Train: RLCEF
Poolside’s key innovation is RLCEF: Reinforcement Learning with Code Execution Feedback. This is fundamentally different from how most labs train coding models.
Traditional Approach (RLHF)
Most models learn from human feedback:
- Model generates code
- Human rates whether the code looks correct
- Model updates based on human preference
The problem: humans are bad at judging code correctness by reading it. We miss edge cases, misunderstand logic, and prefer “clean-looking” code over actually-working code.
Poolside’s Approach (RLCEF)
Poolside trains from execution results:
- Model generates code
- Code actually runs in a real environment
- Tests execute and produce pass/fail signals
- Model updates based on whether the code actually worked
The advantage: the model learns to write code that passes tests, not code that impresses reviewers. This is why Laguna excels on execution-based benchmarks (SWE-bench, DeepSWE) where the code has to actually run correctly.
409K Training Environments
This number is staggering. Poolside built 409,000 distinct training environments where models practice writing code. Each environment is a realistic software engineering scenario:
- Real repositories with real test suites
- Diverse languages, frameworks, and toolchains
- Bug fixing, feature implementation, refactoring
- Different project scales from single files to large codebases
The diversity matters. A model trained on 100 environments learns to solve those 100 types of problems. A model trained on 409K environments learns generalizable coding skills.
This is also why Laguna S 2.1 scores 78.5% on SWE-bench Multilingual. With 409K environments spanning many languages and frameworks, the model develops genuine polyglot capability rather than Python-dominant skills.
Multi-harness Rollouts
During training, Laguna does not just try one solution per problem. It attempts multiple approaches:
- Different algorithms for the same problem
- Various implementation strategies
- Multiple debugging approaches when something fails
The model learns from all attempts, including failures. This teaches:
- When to try a different approach vs. iterate on the current one
- How to recover from wrong starts
- Multiple valid solutions to the same problem
This directly explains Laguna’s strong DeepSWE performance (40.4%). DeepSWE requires exactly this kind of adaptive, multi-attempt reasoning. Models that only learned from single-shot successes struggle when their first approach does not work.
Why Open Weights Matter
Poolside releases model weights under OpenMDW-1.1, which allows commercial use. This is a strategic choice with major implications:
For developers:
- Self-host for zero API cost
- Deploy air-gapped for sensitive codebases
- No vendor lock-in or rate limits
- Audit the model’s capabilities directly
For the ecosystem:
- Fine-tuning for specific domains or languages
- Research and academic use
- Community-driven improvements
- Competitive pressure on closed providers
For the industry:
- A Western alternative to DeepSeek’s open-weight dominance
- Proof that open-weight can match or beat closed models on coding
- More competition drives better models for everyone
Before Poolside, the narrative was: open-weight coding models come from China (DeepSeek, Kimi, Tencent). Western labs keep their best models closed. Poolside breaks that pattern.
The Case Studies: What the Model Can Actually Do
Poolside published three demonstrations that go beyond benchmarks:
Browser Engine From Scratch
Laguna S 2.1 built a working browser rendering engine in 50 minutes across 181 steps. This demonstrates:
- Sustained multi-step execution without losing context
- Architectural decision-making at system level
- Integration of multiple complex components
Self-improvement: Training Harness Optimization
Pointed at Poolside’s own infrastructure, Laguna found optimizations that made the training harness 5.2% faster while using 70% less memory. The model improving its own training pipeline is a powerful demonstration of practical utility.
Mathematical Reasoning: Erdos Problem #397
This 50+ year open problem in combinatorial mathematics was re-derived by Laguna through chain-of-thought reasoning. Not memorized. Derived. This shows the model can reason about novel problems, not just pattern-match against training data.
How Poolside Compares to Other Labs
| Company | Focus | Best Code Model | Open Weights? |
|---|---|---|---|
| Poolside | Code-first | Laguna S 2.1 (70.2% T-Bench) | Yes |
| DeepSeek | General + Code | V4 Pro Max (64.0% T-Bench) | Partial |
| Anthropic | General | Claude Fable 5 (88.0% T-Bench) | No |
| Meta | General | Muse Spark 1.1 (80.0% T-Bench) | Partial |
| Moonshot AI | General + Code | Kimi K3 (88.3% T-Bench) | Yes |
| Tencent | General + Code | Hy3 (71.7% T-Bench) | Yes |
Poolside is unique in being code-focused, open-weight, and Western-based. They sacrifice generality for coding excellence, and the benchmarks show it works.
For detailed comparisons: Laguna vs DeepSeek V4, Laguna vs Kimi K3, Laguna vs Tencent Hy3.
What Comes Next
Poolside trained S 2.1 in under 9 weeks. That rapid iteration cycle means we should expect:
- Larger MoE variants with more active parameters
- Continued improvement of the RLCEF training pipeline
- More training environments (beyond 409K)
- Expanded pool CLI capabilities
- Potential general-purpose models built on RLCEF foundations
The $3B valuation and dedicated compute (4,096 H200s) give Poolside resources to iterate quickly. If RLCEF continues to show these efficiency gains, future models could close the remaining gap to Kimi K3 and Claude Fable 5 while maintaining the open-weight, cost-effective approach.
Check our best AI coding tools guide and open-source coding models ranking for how Poolside fits into the broader AI development landscape.
FAQ
Is Poolside just for coding?
Yes, currently. Their entire product line (Laguna S 2.1, XS 2.1, pool CLI) focuses on software engineering. They do not offer general-purpose chat models, image generation, or other AI capabilities. This specialization is their advantage.
What does the OpenMDW-1.1 license allow?
Commercial use is permitted. You can build and sell products using Laguna S 2.1 without licensing fees. Check the full license text on HuggingFace for specific terms, but the intent is permissive commercial use.
How is Poolside funded and who backs them?
Poolside is valued at $3 billion. While specific investor details are not the focus here, the valuation indicates significant institutional backing. Their 4,096 H200 GPU cluster represents substantial capital investment in training infrastructure.
Can Poolside’s approach work for non-code AI?
RLCEF specifically requires execution feedback, which is natural for code (run tests, see results). Adapting it to other domains (writing, analysis, reasoning) would require defining equivalent “execution” signals. It is possible but not straightforward. For now, Poolside is focused on code where RLCEF has a natural advantage.
How does Poolside compare to DeepSeek as an open-weight provider?
Both release open weights. DeepSeek is larger (1.6T total) but Laguna S 2.1 beats it on coding benchmarks despite being 14x smaller. DeepSeek offers broader model capabilities (general chat, math, vision). Poolside is coding-specialized. For software engineering specifically, Poolside currently wins on efficiency and performance.