Poolside Releases Laguna S 2.1, an Open-Weight Agentic Coding Model Punching Above Its Weight Class on SWE-Bench Multilingual
Poolside has launched Laguna S 2.1, a 118B-parameter open-weight mannequin constructed for agentic coding. It is a Mixture-of-Experts (MoE) mannequin with 8B activated parameters per token. It helps a context window of as much as 1M tokens in each pondering and no-thinking modes. The weights are on Hugging Face beneath an OpenMDW-1.1 license, and the mannequin is sufficiently small to run on a single NVIDIA DGX Spark.
On long-horizon coding benchmarks, Laguna S 2.1 holds its personal towards fashions a number of occasions its dimension, together with DeepSeek-V4-Pro-Max, NVIDIA’s Nemotron 3 Ultra, and Thinking Machines’ Inkling. Laguna S 2.1 is a scale-up of the Laguna XS household, skilled on the identical pre-training information as XS 2.1.
What is Laguna S 2.1
The mannequin prompts roughly 6.8% of its parameters on any given token. All 118B parameters stay resident in reminiscence, however solely ~8B route by means of the community per step. That sparsity is why a mid-size mannequin can behave like a bigger one whereas staying low-cost to serve.
Poolside group publishes weights in BF16, FP8, INT4, and NVFP4, together with official GGUF and MLX conversions and DFlash draft fashions. It went from the beginning of coaching to launch in beneath 9 weeks. Pre-training started on 22 May 2026 on 4,096 NVIDIA H200 GPUs. It is the primary Poolside mannequin the place reinforcement studying ran in FP8 precision.
Performance
Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1 with pondering enabled. That locations it first amongst open, disclosed-size fashions on Poolside’s compiled leaderboard, behind solely bigger or closed techniques. On SWE-Bench Multilingual it scores 78.5%, topping the revealed desk outright. The full comparability Poolside launched is beneath.
| Benchmark | Laguna S 2.1 (118B-A8B) | Tencent Hy3 (295B-A21B) | Inkling (975B-A41B) | Nemotron 3 Ultra (550B-A55B) | DeepSeek-V4-Pro-Max (1.6T-A49B) | Kimi K3 (2.8T-A50B) | Qwen 3.7 Max | Muse Spark 1.1 | Claude Fable 5 |
|---|---|---|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 70.2 | 71.7 | 63.8 | 56.4 | 64.0 | 88.3 | 74.5 | 80.0 | 88.0 |
| SWE-Bench Multilingual | 78.5 | 75.8 | – | 67.7 | 76.2 | – | 78.3 | – | – |
| SWE-Bench Pro (Public) | 59.4 | 57.9 | 54.3 | – | 55.4 | – | 60.6 | 61.5 | 80.3 |
| DeepSWE v1.1 | 40.4 | – | – | – | 9.0 | 69.0 | – | 53.3 | 70.0 |
| SWE Atlas (Codebase QnA) | 46.2 | – | – | – | 27.2 | – | – | 42.2 | – |
| Toolathlon Verified | 49.7 | – | 45.5 | 34.3 | 55.9 | – | – | 75.6 | – |
The clearest sign is DeepSWE v1.1, which nonetheless has actual headroom. There, Laguna S 2.1 scores 40.4% towards DeepSeek-V4-Pro-Max’s 9.0%, with roughly one-sixth the energetic parameters. Closed frontier fashions resembling Claude Fable 5 and Kimi K3 nonetheless lead on a number of benchmarks. Poolside’s declare is concerning the weight class, not the outright prime. Trajectories from the ultimate analysis set is revealed at trajectories.poolside.ai.
Two pondering modes, and the place the rating comes from
Laguna S 2.1 has two modes: off and max, with max enabled by default. In max mode the mannequin units its personal test-time compute funds. Poolside is delivery with out user-configurable low/medium/excessive effort management for now.
Max pondering lifts Terminal-Bench 2.1 from 60.4% to 70.2%. It lifts DeepSWE from 16.5% to 40.4%. Those positive factors price tokens: DeepSWE trajectories run about 249k completion tokens in pondering mode towards 99k with out. Poolside group experiences coherent, productive reasoning operating for hours and tons of of 1000’s of tokens.
Three revealed runs
Poolside group shared three unedited trajectories to indicate conduct relatively than scores. In one, the mannequin constructed a working HTML/CSS browser engine from an empty folder. That run took 181 steps throughout a 50-minute session, with no human intervention, and it validated its personal output towards headless Chromium.
In a second, the mannequin optimized Poolside’s personal agent harness. It made the harness 5.2% quicker with roughly 71% decrease reminiscence allocation, changing O(n²) string concatenation with buffers. In a 3rd, it re-derived Erdős Problem #397 offline in Perl over 68 minutes, because the sandbox had no Python. That result’s an impartial rediscovery; GPT-5.2 Pro solved the identical downside in January 2026, and Laguna’s data cutoff is November 2025.
Can you truly deploy it?
Sizing makes use of the complete 118B parameters, not the 8B energetic depend, as a result of each skilled stays in reminiscence. At 4-bit (NVFP4 or INT4) the weights want about 59 GB, which inserts comfortably on a single DGX Spark’s 128 GB of unified reminiscence. At FP8 they want about 118 GB, nonetheless inside a single Spark or a single H200. At BF16 they want about 236 GB, which calls for 2 linked Sparks or a multi-GPU datacenter node.
Poolside labored with NVIDIA to optimize inference from TRT-LLM serving to NVFP4 on Blackwell, all the way down to a single DGX Spark. It shipped day-one help for vLLM, SGLang, and Ollama. Hosted entry runs by means of OpenRouter, free at 256K context and paid on the full 1M context for $0.10 / $0.20 / $0.01 per 1M enter / output / cache-read tokens. The mannequin can be on Baseten, Kilo, Prime Intellect Prime Lab, and ZML.
Key Takeaways
- Laguna S 2.1 is a 118B-total / 8B-active MoE coding mannequin with a 1M-token context, open beneath OpenMDW-1.1.
- It scores 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-Bench Multilingual, main open disclosed-size fashions.
- At 4-bit it runs on a single NVIDIA DGX Spark; FP8 suits one Spark or H200, BF16 wants two.
- Default ‘max pondering’ drives a lot of the rating, at an actual token price (DeepSWE: 16.5% → 40.4%).
- Trained in beneath 9 weeks on 4,096 H200 GPUs, it’s Poolside’s third mannequin in beneath three months.
Sources: Poolside Technical— Introducing Laguna S 2.1, Robert McHardy on X, Poolside on X and Hugging Face model card.
The submit Poolside Releases Laguna S 2.1, an Open-Weight Agentic Coding Model Punching Above Its Weight Class on SWE-Bench Multilingual appeared first on MarkTechPost.
