Poolside Laguna S 2.1: The West's Most Capable Open-Weight Coding Model (2026)
Poolside Laguna S 2.1: The West’s Most Capable Open-Weight Coding Model (2026)
Poolside just dropped the model that changes the entire open-weight conversation. Laguna S 2.1 is a 118B Mixture of Experts model that only activates 8B parameters per token, yet it beats DeepSeek V4 Pro Max (a 1.6 trillion parameter model) on both SWE-bench Pro and DeepSWE. It has a 1 million token context window. It is free on OpenRouter at 256K context. And the weights are fully open.
This is not incremental progress. This is an 8B-active model outperforming competitors that are 5x, 6x, even 14x its active size. The West finally has an open-weight coding model that can compete with the best from China, and it is punching absurdly far above its weight class.
What Is Laguna S 2.1?
Laguna S 2.1 is Poolside’s flagship code model, released July 21, 2026. Here are the raw specs:
| Spec | Value |
|---|---|
| Architecture | Mixture of Experts (MoE) |
| Total parameters | 118B |
| Active per token | 8B |
| Context window | 1,000,000 tokens |
| Training hardware | 4,096 H200 GPUs |
| Training time | Under 9 weeks |
| License | OpenMDW-1.1 |
| Weights | HuggingFace, Ollama, vLLM, SGLang |
| API (256K) | Free on OpenRouter |
| API (1M) | $0.10/$0.20/$0.01 per 1M tokens |
The MoE architecture is the key. With only 8B parameters activated per token, Laguna S 2.1 runs dramatically faster and cheaper than dense models of equivalent capability. You get the intelligence of something much bigger with the inference cost of something small.
Benchmark Performance
The numbers speak for themselves. Laguna S 2.1 dominates its weight class and punches well into territory occupied by models many times its size.
| Benchmark | Laguna S 2.1 | Notes |
|---|---|---|
| Terminal-Bench 2.1 | 70.2% | #1 in weight class |
| Terminal-Bench (no thinking) | 60.4% | Still competitive |
| SWE-bench Pro | 59.4% | Beats DeepSeek V4 (55.4%) |
| SWE-bench Multilingual | 78.5% | Excellent polyglot |
| DeepSWE | 40.4% | Beats DeepSeek V4 (9.0%) |
For context on what these numbers mean against the wider field: Tencent Hy3 scores 71.7% on Terminal-Bench but uses 21B active params (2.5x more). Kimi K3 hits 88.3% but is a 2.8 trillion parameter model (24x larger). The efficiency story here is extraordinary.
The model that Laguna S 2.1 embarrasses most directly is DeepSeek V4 Pro Max. That model has 1.6T total parameters and 49B active per token. Laguna S 2.1, with its 8B active params, beats it on SWE-bench Pro (59.4 vs 55.4) and absolutely destroys it on DeepSWE (40.4 vs 9.0). Read the full comparison for details.
Thinking Mode: The Secret Weapon
Laguna S 2.1 ships with both thinking and non-thinking modes, and the difference is massive.
- Terminal-Bench: 60.4% without thinking, 70.2% with thinking (+16.2% relative improvement)
- DeepSWE: 16.5% without thinking, 40.4% with thinking (+145% relative improvement)
That DeepSWE jump is not a typo. Thinking mode nearly triples the model’s score on the hardest agentic coding benchmark. For complex multi-file tasks that require planning, you absolutely want thinking enabled. Read the full thinking mode guide for configuration details.
Real-World Case Studies
Benchmarks tell one story. What the model actually does in practice tells another. Poolside published three case studies that demonstrate capability beyond what benchmarks capture:
Built a Browser Engine From Scratch
Laguna S 2.1 constructed a working browser rendering engine in 50 minutes across 181 steps. Not a wrapper around WebKit. A browser engine. From nothing. This is the kind of sustained, multi-step reasoning that separates real coding models from benchmark optimizers.
Optimized Its Own Training Harness
The model was pointed at Poolside’s own training infrastructure and found optimizations that made the harness 5.2% faster while using 70% less memory. A model that can improve its own training pipeline is a different category of useful.
Re-derived Erdos Problem #397
This is a 50+ year open problem in combinatorial mathematics. Laguna S 2.1 independently re-derived a known result for this problem. Not by memorizing it from training data, but through chain-of-thought reasoning. This demonstrates genuine mathematical reasoning capability, not just code completion.
Pricing and Access
Poolside made access straightforward:
Free tier (OpenRouter): Up to 256K context, no cost. This is enough for most coding tasks.
Paid tier (1M context): $0.10 per million input tokens, $0.20 per million output tokens, $0.01 per million cached tokens. For comparison, Claude Sonnet 5 costs $2/$10, and Grok 4.5 costs $2/$6. Laguna S 2.1 is 20x cheaper than Sonnet on input and 50x cheaper on output.
Self-hosted: Download weights from HuggingFace and run via Ollama, vLLM, or SGLang. Zero API costs. See our local deployment guide for hardware requirements.
Check our AI API pricing comparison for the full picture across all providers.
The Pool CLI Agent
Poolside ships their own terminal agent called pool. Think Claude Code but powered by Laguna S 2.1. It connects directly to the model’s 1M context window and thinking capabilities, giving you a first-party agentic coding experience without third-party wrappers.
If you already use Aider or similar tools, pool offers a tighter integration with the model’s native capabilities. Full details in our Pool CLI guide.
How It Was Trained
Poolside’s training approach is genuinely novel:
- RLCEF (Reinforcement Learning with Code Execution Feedback): Instead of human preferences, the model learns from actual code execution results.
- 409K training environments: Not toy problems. Real, diverse software engineering scenarios.
- Multi-harness rollouts: The model attempts problems multiple ways and learns from all outcomes.
- 9 weeks on 4,096 H200 GPUs: Fast iteration cycles let Poolside ship improvements rapidly.
This training approach explains why the model excels at real-world coding tasks (SWE-bench, Terminal-Bench) rather than just isolated code generation.
Laguna XS 2.1: The Smaller Sibling
If you want to run Poolside locally on consumer hardware, Laguna XS 2.1 is the option. Released July 2, 2026:
| Spec | XS 2.1 |
|---|---|
| Total params | 33B |
| Active per token | 3B |
| Context | 262K |
| SWE-bench Multilingual | 63.1% |
| Terminal-Bench | 33.4% |
| Price | Free on OpenRouter |
With only 3B active parameters, XS 2.1 runs comfortably on machines with 24GB VRAM. See our VRAM requirements guide for details on what hardware you need.
Who Should Use Laguna S 2.1?
Use it if you want:
- The best open-weight coding model available in the West
- Free or near-free API access for coding tasks
- Self-hosting capability with no vendor lock-in
- A model that handles million-token contexts for large codebases
- Performance that beats proprietary models costing 20-50x more
Consider alternatives if you need:
- Absolute peak performance regardless of cost (Claude Fable 5 at 88.0% T-Bench)
- The fastest inference speed (smaller models or purpose-built fast providers)
- Non-coding tasks (general models may be better)
For most developers building with AI coding tools, Laguna S 2.1 represents the best value proposition available today. It competes with models that cost orders of magnitude more, and you can run it yourself. That combination did not exist before this release.
Check our best open-source coding models ranking and best AI coding tools guide for how Laguna S 2.1 fits into the broader landscape.
FAQ
Is Laguna S 2.1 really free?
Yes, on OpenRouter at up to 256K context. The 1M context tier costs $0.10/$0.20 per million tokens, which is still extremely cheap. You can also download the weights and run it yourself at zero ongoing cost.
How does Laguna S 2.1 compare to Claude Sonnet 5?
Laguna S 2.1 (59.4% SWE-bench Pro) is close to Sonnet 5 (63.2%) while being free or 20-50x cheaper on API. For most coding tasks the gap is negligible, and self-hosting gives you unlimited use. See the full comparison.
Can I run Laguna S 2.1 on my own hardware?
The full S 2.1 (118B total) requires serious hardware for local deployment. For consumer-grade machines, use Laguna XS 2.1 (33B total, 3B active) which fits on a single GPU with 24GB VRAM. Read the local deployment guide.
What license does Laguna S 2.1 use?
OpenMDW-1.1, which allows commercial use. You can build products on top of Laguna S 2.1 without licensing fees.
Should I use thinking mode or non-thinking mode?
For simple, fast tasks (autocomplete, short functions), non-thinking mode is faster and still scores 60.4% on Terminal-Bench. For complex multi-file tasks, always use thinking mode since the jump to 70.2% T-Bench and 40.4% DeepSWE is enormous. More details in the thinking mode guide.