Poolside's Laguna S 2.1 matches or beats far larger models on coding benchmarks with open weights
- Poolside released Laguna S 2.1, a 118B total parameter Mixture-of-Experts model with only 8B activated parameters per token and a context window up to 1M tokens, trained start to launch in under nine weeks.
- On Terminal-Bench 2.1, Laguna S 2.1 scores 70.2%, beating much larger models like DeepSeek-V4-Pro-Max (1.6T-A49B, 64.0%) and Inkling (975B-A41B, 63.8%) despite being roughly 13x smaller.
- The model also scores 78.5% on SWE-Bench Multilingual and 59.4% on SWE-Bench Pro, both ahead of or on par with much bigger closed and open models including Tencent Hy3 (295B-A21B) and Nemotron 3 Ultra (550B-A55B).
- Poolside released full evaluation trajectories for every trial at trajectories.poolside.ai and also shipped a smaller Laguna XS 2.1 (33B-A3B) model alongside quantized GGUF and DFlash speculative decoding drafters for local hardware like Strix Halo and DGX Spark.
- Hacker News commenters found that thinking mode is disabled by default in some deployment recipes (including OpenRouter), which silently caps output and makes the model look worse until enable_thinking and max_new_tokens are fixed.
Hacker News 의견들
Open weights and similar performance to DeepSeek V4 and Inkling at the size of Nemotron 3 Super, that's insane if it holds up.
How did they even pull this off? Their own writeup says they didn't add intelligence, they improved behaviors like more verification, less assuming, not declaring victory early, and being more persistent. Also nine weeks start to launch is wild.
This thing is going to run great on Strix Halo and DGX Spark boxes. Already saw Q4_K_M and a DFlash speculative decoding drafter on the HF release, hoping every model ships with a dflash drafter from now on.
This fills the exact gap that's been missing: self hosted, good enough, MoE so it's fast on limited bandwidth setups like Strix Halo. Nothing beat my dual 32GB desktop running Gemma 4 or Qwen 3.6 dense until now.
Been using it for 3 hours and I'm probably switching from Claude Code to this.
118B total, 8B active MoE, long context, open weights, this is a real sweet spot for size vs performance. Never heard of this lab before but very excited to try it.
Testing now, at least competitive with DS4-Flash. It found things in my C codebase that only gpt-5.2 caught before, but also hallucinated a wrong memfd_create/mmap claim that even Sol's own review missed until I pointed it out.
This is the launch of the day, crushing Google's releases. First US model that's actually competitive with DeepSeek V4 Flash on pricing too.
We already got a usable PR out of it on our repo, this model is not a joke.
Would love a quant for us 64GB paupers. Someone's already working on a GGUF on HuggingFace though.
Q4_K_M is 75GB, right at my 64GB limit, I wouldn't quantize further, better to do partial weight residency and stream from SSD instead.
Heads up: a ton of people testing this are getting disappointing benchmaxxed looking results because THINKING isn't enabled by default. You need to pass enable_thinking true and bump max_new_tokens past 32k or it cuts off the reasoning entirely. Once I fixed that the code quality jumped a lot.
Even OpenRouter's official provider has this thinking-disabled bug, hope it's an easy fix on their end.
This 128B model beating 1.6T DeepSeek V4 on most coding benchmarks is a strong signal. Also I like that Poolside compares against much bigger models like 2.5T Kimi K3 instead of just cherry picking its own weight class.