Poolside Releases Laguna S 2.1, an Open-Weight Agentic Coding Model Punching Above Its Weight Class on SWE-Bench Multilingual
Poolside released Laguna S 2.1, an 118B-parameter Mixture-of-Experts (MoE) coding model with only 8B active parameters per token, enabling high performance on limited hardware. The model achieves state-of-the-art results among open, disclosed-size models, scoring 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-Bench Multilingual. It features a 1M-token context window and a "max thinking" mode that significantly boosts performance on complex tasks like DeepSWE (from 16.5% to 40.4%) at the cost of hi
Analysis
TL;DR
- Poolside released Laguna S 2.1, an 118B-parameter Mixture-of-Experts (MoE) coding model with only 8B active parameters per token, enabling high performance on limited hardware.
- The model achieves state-of-the-art results among open, disclosed-size models, scoring 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-Bench Multilingual.
- It features a 1M-token context window and a "max thinking" mode that significantly boosts performance on complex tasks like DeepSWE (from 16.5% to 40.4%) at the cost of higher token usage.
- Weights are available in multiple precisions (BF16, FP8, INT4, NVFP4), allowing deployment on a single NVIDIA DGX Spark (128GB) when quantized to 4-bit.
- Training was completed in under nine weeks on 4,096 NVIDIA H200 GPUs, marking it as Poolside's third major release in less than three months.
Why It Matters
Laguna S 2.1 demonstrates that efficient MoE architectures can compete with significantly larger dense or hybrid models, challenging the assumption that sheer parameter count is the primary driver of coding capability. Its ability to run on consumer-grade or edge-like hardware (single DGX Spark) lowers the barrier for private, secure agentic coding deployments. Furthermore, the emphasis on long-horizon "thinking" modes highlights a shift toward models capable of sustained, multi-step reasoning rather than just immediate code generation.
Technical Details
- Architecture: 118B total parameters with 8B activated per token (approx. 6.8% sparsity). Uses a Mixture-of-Experts design where all experts remain in memory during inference.
- Context & Modes: Supports up to 1M tokens context. Features two thinking modes: "off" and "max" (default). The "max" mode allows the model to self-determine test-time compute budgets, leading to longer reasoning chains.
- Training & Precision: Pre-trained on 4,096 NVIDIA H200 GPUs in under nine weeks. Notably, this is the first Poolside model to use FP8 precision for reinforcement learning. Weights are provided in BF16, FP8, INT4, and NVFP4 formats.
- Performance Benchmarks:
- Terminal-Bench 2.1: 70.2% (with thinking).
- SWE-Bench Multilingual: 78.5%.
- DeepSWE v1.1: 40.4% (a significant jump from 16.5% without thinking).
- Deployment: Optimized for vLLM, SGLang, and Ollama. Requires ~59GB VRAM for 4-bit (fits DGX Spark), ~118GB for FP8, and ~236GB for BF16.
Industry Insight
- Efficiency over Scale: The success of an 8B-active parameter model competing with models having tens or hundreds of billions of active parameters suggests that architectural efficiency (MoE) and training data quality are becoming more critical than raw model size for specific domains like coding.
- Agentic Reasoning Costs: The substantial performance gain from "thinking" modes comes with a heavy token cost (e.g., DeepSWE trajectories increasing from 99k to 249k tokens). Practitioners must balance latency/cost against accuracy when deploying agentic workflows.
- Hardware Democratization: By supporting NVFP4 and fitting into 128GB unified memory systems, Poolside enables organizations to deploy powerful coding agents locally or on smaller clusters, reducing reliance on massive cloud GPU instances for inference.
Disclaimer: The above content is generated by AI and is for reference only.