Today we're releasing Laguna S 2.1, our most capable model to date....
Poolside@poolsideai
62 views
Jul 22, 2026
~3 min read
Advertisement
1
Today we're releasing Laguna S 2.1, our most capable model to date.
It's a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and thinking and no-thinking modes.
Capable enough to hold its own against models many times its size. Small enough to run on a single @NVIDIAAI DGX Spark.
Laguna S 2.1 is fully open under OpenMDW-1.1, with weights available today on @huggingface
poolside.ai/blog/introduci…
It's a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and thinking and no-thinking modes.
Capable enough to hold its own against models many times its size. Small enough to run on a single @NVIDIAAI DGX Spark.
Laguna S 2.1 is fully open under OpenMDW-1.1, with weights available today on @huggingface
poolside.ai/blog/introduci…
0:07
2
Laguna S 2.1 is, as far as we can measure, the most capable agentic coding model in its weight class.
On Terminal-Bench 2.1 it scores 70.2, sitting beside models 5–25x its size and ahead of several of them.
And on DeepSWE from @datacurve, the hardest long-horizon benchmark we ran, Laguna S 2.1 scores 40.4, outperforming some open models with more than 1T parameters.
For every score we publish today, we're releasing the full trajectory of every trial in the final evaluation set at trajectories.poolside.ai
On Terminal-Bench 2.1 it scores 70.2, sitting beside models 5–25x its size and ahead of several of them.
And on DeepSWE from @datacurve, the hardest long-horizon benchmark we ran, Laguna S 2.1 scores 40.4, outperforming some open models with more than 1T parameters.
For every score we publish today, we're releasing the full trajectory of every trial in the final evaluation set at trajectories.poolside.ai
3
Laguna S 2.1 is built to run locally.
We heard the ask for a serious open model in the 80–120B range. Laguna S 2.1 is our answer, and we want it to redefine what this size class can do.
Weights ship in BF16, FP8, INT4 and NVFP4, with official GGUF and MLX conversions, plus official DFlash draft models for faster inference. Day-one support across llama.cpp, @ollama, @vLLM, @sgl_project, NVIDIA TRT-LLM, @huggingface transformers and @zml_ai.
Watch it run on an @NVIDIAAI DGX Spark.
We heard the ask for a serious open model in the 80–120B range. Laguna S 2.1 is our answer, and we want it to redefine what this size class can do.
Weights ship in BF16, FP8, INT4 and NVFP4, with official GGUF and MLX conversions, plus official DFlash draft models for faster inference. Day-one support across llama.cpp, @ollama, @vLLM, @sgl_project, NVIDIA TRT-LLM, @huggingface transformers and @zml_ai.
Watch it run on an @NVIDIAAI DGX Spark.
0:39
4
What excites us most is the model behavior.
Laguna S 2.1 verifies its work instead of taking things for granted. It doesn't declare victory early. And it's remarkably persistent.
We pointed it at our own agent harness in an automated loop and after ~690k tokens, it had made the harness 5.2% faster while cutting memory allocation by ~70%. When the remaining gains became too small for the benchmark to resolve, it didn't declare victory or fake progress. It moved to the metric it could still measure and kept driving memory down.
Watch the full run, turn by turn: trajectories.poolside.ai
Laguna S 2.1 verifies its work instead of taking things for granted. It doesn't declare victory early. And it's remarkably persistent.
We pointed it at our own agent harness in an automated loop and after ~690k tokens, it had made the harness 5.2% faster while cutting memory allocation by ~70%. When the remaining gains became too small for the benchmark to resolve, it didn't declare victory or fake progress. It moved to the metric it could still measure and kept driving memory down.
Watch the full run, turn by turn: trajectories.poolside.ai
5
Laguna S 2.1 is a scale-up of the Laguna XS family, trained on exactly the same pre-training data as Laguna XS 2.1. The step was scale, training-code fixes and small recipe changes.
Most of the capability gain came from post-training: SFT, partly on synthetic data, followed by RL on tasks the model could not yet solve reliably. We used longer rollout budgets, improved sandbox infrastructure and multi-harness rollouts so the learned behavior transfers across agent scaffolds. It’s also our first model to run RL in FP8.
Laguna S 2.1 is another proof that the Model Factory is working. With every training run, we improve the data, infrastructure, evaluations and RL systems behind the next one, helping us release more capable models faster.
Most of the capability gain came from post-training: SFT, partly on synthetic data, followed by RL on tasks the model could not yet solve reliably. We used longer rollout budgets, improved sandbox infrastructure and multi-harness rollouts so the learned behavior transfers across agent scaffolds. It’s also our first model to run RL in FP8.
Laguna S 2.1 is another proof that the Model Factory is working. With every training run, we improve the data, infrastructure, evaluations and RL systems behind the next one, helping us release more capable models faster.
6
As always, there's a lot of work ahead. We're releasing Laguna S 2.1 with known limitations that we're already addressing in the next iteration:
- It can overthink, especially on hard math. There's currently no setting between no-thinking and max. Intermediate effort control is at the top of the roadmap.
- In third-party harnesses, it sometimes trusts its memory of a tool's interface on first use instead of the schema in front of it. A rejected call and retry usually corrects it.
- Tool arguments expecting JSON arrays can be incorrectly escaped.
Tell us what breaks on Discord: discord.gg/ZGv44HRgy
- It can overthink, especially on hard math. There's currently no setting between no-thinking and max. Intermediate effort control is at the top of the roadmap.
- In third-party harnesses, it sometimes trusts its memory of a tool's interface on first use instead of the schema in front of it. A rejected call and retry usually corrects it.
- Tool arguments expecting JSON arrays can be incorrectly escaped.
Tell us what breaks on Discord: discord.gg/ZGv44HRgy
7
Laguna S 2.1 is available today, everywhere you build.
Weights are live on @huggingface under OpenMDW-1.1, in BF16, FP8, INT4 and NVFP4, with official GGUF, MLX and DFlash conversions.
Use it through @OpenRouter or @vercel’s AI Gateway, deploy it with @basetenco, or build with it directly in your favourite harness @kilocode, @NousResearch @cline, @pidotdev or @openclaw
Run it locally with @ollama @vllm_project, @SGLang or llama.cpp; serve it with NVIDIA TRT-LLM on a single @NVIDIAAI DGX Spark; or run it on almost any hardware through @zml_ai LLMD.
Laguna S 2.1 is also available for post-training through @PrimeIntellect Prime Lab.
The best way to try it is with pool, our terminal coding agent: poolside.ai/get-started
For the simplest way to chat with the model, head to chat.poolside.ai, no login required
Go build something cool!
Weights are live on @huggingface under OpenMDW-1.1, in BF16, FP8, INT4 and NVFP4, with official GGUF, MLX and DFlash conversions.
Use it through @OpenRouter or @vercel’s AI Gateway, deploy it with @basetenco, or build with it directly in your favourite harness @kilocode, @NousResearch @cline, @pidotdev or @openclaw
Run it locally with @ollama @vllm_project, @SGLang or llama.cpp; serve it with NVIDIA TRT-LLM on a single @NVIDIAAI DGX Spark; or run it on almost any hardware through @zml_ai LLMD.
Laguna S 2.1 is also available for post-training through @PrimeIntellect Prime Lab.
The best way to try it is with pool, our terminal coding agent: poolside.ai/get-started
For the simplest way to chat with the model, head to chat.poolside.ai, no login required
Go build something cool!
