By Shakti Tiwari — Nifty Option Trader, Research Analyst & XGBoost Expert
NISM Series XII Certified (Registration: NISM-202500241080)
Poolside AI recently released Lagona S2.1, and for a mid-size US lab it is genuinely impressive. It is not trying to be another flashy vibe-coding model — it is a practical, locally runnable coding model built for developers who want strong real-world performance without a massive multi-GPU server. This breakdown covers the architecture, benchmark scores, local inference speeds, and exactly how to run it.
What Is Lagona S2.1?
Lagona S2.1 is a 118-billion-parameter mixture-of-experts (MoE) model with just 8 billion active parameters per token. It supports a 1-million-token context window and ships in both thinking and non-thinking modes.
The headline number: it went from the start of training to launch in under 9 weeks. But the more important story is efficiency — this is a model you can actually own and run.
Why a 3090 Can Host a 118B Model
Because only ~8B parameters are active per token, the model fits on accessible consumer hardware:
- Runs on a single Nvidia DGX Spark
- Runs on a single RTX 3090 (24 GB) via native NV4 quantization
- Fits on 128 GB MacBook (the smaller XS 2.1 variant)
That is a level of deployability most frontier labs cannot offer at this parameter count.
Benchmark Scores (Terminal Bench 2.1 / SWE-bench)
On the World of AI vibe-coding benchmark, Lagona S2.1 ranks 6th overall in the open-weight category — competing directly with a 753B-parameter model while staying locally hostable.
| Benchmark | Score |
|---|---|
| SWE-bench Multilingual | 78.5% |
| Terminal Bench 2.1 | 70.2% |
| SWE-bench Pro | 59.9% |
| Deep Suite | 40.4% |
On Terminal Bench 2.1 / Deep SWE it ranks #11 overall, outperforming Gemini 3.5 Flash, Kimi K2.7 Code, and Claude Sonnet — remarkable for a 118B open-weight MoE.
Coding Quality vs Speed
Lagona XS 2.1 vs Qwen 3.6 (35B) — both built a playable Tetris in one HTML file over 3 iterations:
- Lagona XS 2.1: ~45K tokens at ~158 tok/s
- Qwen 3.6 35B: ~39K tokens at ~81 tok/s
Nearly identical quality, roughly 2× the speed.
Lagona S2.1 vs GLM 5.2 / HY3 on Doodle Jump, Air Hockey, Geometry Dash:
- Lagona generated just 10.3K tokens vs GLM 5.2's 26.4K and HY3's 10.4K
- Geometry Dash output was arguably the best of the three — despite competing with a 753B model (1/7th the size)
- Only model of the three that could run locally on a 128 GB MacBook
Local Inference Speed (Native NV4 Quantization)
| Hardware | Tokens / sec |
|---|---|
| Nvidia DGX Spark | ~19 tok/s |
| RTX Pro 6000 | ~109 tok/s |
| 4× RTX 5090 | ~146 tok/s |
For a 117B-parameter MoE with 8.5B active params, that is serious performance from hardware you can actually buy.
What It Generates Well
Tested through the World of AI benchmark tool, Lagona S2.1 produced:
- A fully playable FPS shooter with multiple waves and a reload function (on a 96 GB local rig)
- An F1 drifting donut simulation — the first model the reviewer saw nail the tire-smoke drift animation
- A realistic aquarium with lighting, shaders, water texture, and live controls (clarity, current speed, time of day, orbit toggle)
- A functional Mac OS clone with SVG app icons, terminal, calculator, and a snake game
- An animated lava lamp in CSS+SVG with working blob physics
It is not flawless — the SVG butterfly failed to build, and some generations had small visual bugs — but the consistency across domains on local hardware is the headline.
How to Run Lagona S2.1
- Local weights — install the open weights via Hugging Face
- Poolside chatbot — use both model sizes with thinking/non-thinking modes
- OpenRouter — API access at $0.10 / 1M input tokens, $0.20 / 1M output tokens
- World of AI benchmark — run it completely free, with access to the reviewer's prompts
Bottom Line
Lagona S2.1 does not beat frontier models like GPT-5.6, Claude Fable 5, or Kimi K3 outright. What it does is deliver strong real-world coding performance in a model developers can actually run locally — on a 3090, a DGX Spark, or a MacBook. An 118B model trained in under nine weeks that competes with models several times its size, and fits on hardware you own, is one of the most interesting open-weight releases in a long time.
Frequently Asked Questions
What is Poolside Lagona S2.1?
A 118-billion-parameter mixture-of-experts LLM from Poolside AI with 8 billion active parameters per token, a 1-million-token context, and thinking/non-thinking modes. It is built for local, practical coding workflows.
Can Lagona S2.1 run on a single GPU?
Yes. Via native NV4 quantization it runs on a single RTX 3090 (24 GB), an Nvidia DGX Spark, or a 128 GB MacBook — far more accessible than typical 100B+ models.
How fast is Lagona S2.1 locally?
Around 19 tok/s on a DGX Spark, ~109 tok/s on an RTX Pro 6000, and ~146 tok/s across four RTX 5090s. The smaller XS 2.1 hit ~158 tok/s on the Tetris test.
Is Lagona S2.1 free to use?
The open weights are free to run locally. It is also available free through the World of AI benchmark tool, and paid via OpenRouter at $0.10/$0.20 per million input/output tokens.
How does Lagona S2.1 compare to Qwen and GLM?
On coding tests it matched or beat much larger models on quality while using fewer tokens and running 2× faster. It outperformed GLM 5.2 and HY3 on token efficiency and was the only one runnable locally on a 128 GB MacBook.
Disclaimer: Research summary of a third-party model release, not investment or financial advice. Verify specs on the official Poolside / Hugging Face pages before deployment.
Related: Mac Mini M4 Pro for Local AI — 2026 Verdict · My book: Option Trading with AI · NSE research site
💬 Found this useful?
If this helped your trading or ML workflow, drop a comment with your take — and react ❤️ so other NSE traders find it.
I publish NSE research, local-AI model breakdowns, and code-first trading playbooks regularly.
📕 My book: Option Trading with AI (XGBoost, Transformers & Quantized Models for the Retail Nifty Trader) → https://www.amazon.in/dp/B0H9ZNTBPK
🔗 More: NSE research site · X / Twitter
— Shakti Tiwari, Nifty Option Trader & Research Analyst (NISM Series XII)
Top comments (0)