Open Source

Laguna S 2.1 beats Deepseek V4 Pro, costs less than V4 Flash

118B-A8B MoE model tops coding benchmarks at a fraction of the price.

Deep Dive

Finally, the banger from Laguna: S 2.1 (118B-A8B) scores 70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench, 59.4% on multilingual SWE-bench Pro, plus 40.4%, 46.2%, and 49.7% on other benchmarks. The poster says it will probably be great for 64GB+ RAM and VRAM setups.

Key Points
  • Scores 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-bench, outperforming Deepseek V4 Pro.
  • Uses a 118B total parameter MoE with 8B active parameters, optimized for 64GB+ RAM/VRAM.
  • Priced lower than Deepseek V4 Flash while delivering better performance across coding and tool-use tasks.

Why It Matters

Democratizes high-end coding AI by blending superior benchmarks with budget-friendly hardware and pricing.

📬 Get the top 10 AI stories daily