from openai import OpenAI
import os
client = OpenAI(
api_key=os.getenv("POOLSIDE_API_KEY"),
base_url="https://inference.poolside.ai/v1"
)
response = client.chat.completions.create(
model="poolside/laguna-s-2.1",
messages=[{
"role": "user",
"content": "What are channels in Go?"
}]
)Foundation models built for agentic coding.
Poolside trains models from scratch: our own data, our own infrastructure, our own reinforcement learning. Competitive with leading models at a fraction of the compute.
laguna
s2.1laguna
xs2.1
All models:
OpenAI-compatible Chat API
Weights on Hugging Face
Licenses vary by model
These models learned to code inside ❯ pool
All Laguna models are trained with reinforcement learning inside our agent harness; they
perform at their best in pool , or you can use any ACP client.
Frontier-class reasoning at mid-size cost. A 118B total parameter model with 8B activated. Trained in-house
on 30T tokens.
Our strongest model for agentic coding and long-horizon work. Competitive with models several
times its size on Terminal-Bench 2.1 and DeepSWE. Small enough to run on a single NVIDIA DGX
Spark.
Terminal-Bench 2.1
Terminal-Bench 2.1 Resolved tasks on Terminal-Bench 2.1.
SWE-Bench Multilingual
SWE-Bench Multilingual Resolved tasks on SWE-Bench Multilingual.
SWE-Bench Pro (Public Dataset)
SWE-Bench Pro (Public Dataset) Resolved tasks on SWE-Bench Pro (Public Dataset).
DeepSWE
DeepSWE Resolved tasks on DeepSWE.
SWE Atlas (Codebase QnA)
SWE Atlas (Codebase QnA) Resolved tasks on SWE Atlas (Codebase QnA).
Toolathlon Verified
Toolathlon Verified Resolved tasks on Toolathlon Verified.
Benchmarks as of 21 July 2026. pass@1 averaged over 4 attempts per task, except DeepSWE, SWE Atlas (Codebase QnA) and Toolathlon Verified that had 3 attempts per task. For all benchmarks we take the maximum of the vendor self-reported score, benchmark author leaderboard or third-party leaderboard (Artificial Analysis), except SWE Atlas (Codebase QnA) where we do not use third-party leaderboard figures.
Intelligence vs. model size
Our lightest and fastest agentic coding model. A 33B total parameter model with 3B activated. Trained in-house on 15T tokens.
A second-generation architecture and our first open-weight model, built on everything we've
learned since training Laguna across
synthetic data and RL. Laguna performs
at its best in our coding agent, as a strong model for rapid agentic iteration.
m
1
xs
2.1
SWE-bench Verified
SWE-bench Verified Resolved tasks on SWE-bench Verified.
SWE-bench Multilingual
SWE-bench Multilingual Resolved tasks on SWE-bench Multilingual.
SWE-Bench Pro
SWE-Bench Pro Resolved tasks on SWE-Bench Pro.
Terminal-Bench 2.0
Terminal-Bench 2.0 Resolved tasks on Terminal-Bench 2.0.
Benchmarks as of 2 July 2026. † We have chosen to include dense models with larger activated parameter counts to highlight the relative efficiency of MoE models.
Quantize it or serve it. The weights are yours. Bring it into your existing tools and harnesses in minutes.
We believe the West needs strong open-weight models, and we're committed to contributing to that ecosystem. We release the Laguna family under permissive licences from OpenMDW to Apache 2.0, inviting the community to evaluate, fine-tune, and build on our work.
Get started with Laguna. Free to use for a limited time.
$ curl -fsSL https://downloads.poolside.ai/pool/install.sh | sh
poolv1.0.6
• Connected to agent server · Laguna S 2.1
› Audit the checkout flow for race conditions
● Thinking
⎿ Tracing order creation and transaction boundaries
● Read(src/checkout/createOrder.ts)
● Edit(src/checkout/createOrder.ts)
● Bash(pnpm test checkout)
⎿ ✓ 12 tests passed in 284ms
⠹ Working… (12s)
Laguna in your terminal with pool.
$ docker run --gpus all poolside/laguna
NVIDIA DGX Spark · GB10 Grace Blackwell
128 GB unified memory · CUDA device ready
Loading poolside/laguna-s-2.1
Precision FP4 · Context 131,072
Starting OpenAI-compatible inference server...
✓ Ready at http://localhost:8000/v1
Run Laguna locally.
Work with Poolside Desktop Assistant.
How we build these models. The Model Factory, our training infrastructure, and everything we publish as we go.
Research at PoolsideFootnotes.
All benchmarking for Laguna S 2.1 and Laguna XS 2.1 was completed using the Laude Institute's Harbor Framework with our agent harness, a maximum of 500 steps and sandboxed execution via our internal sandbox service. We report mean pass@1 averaged over multiple attempts per task. Final run trajectories for Laguna S 2.1 evaluation are available to view and download at trajectories.poolside.ai.
Some base task images and verifiers were patched to fix infrastructure reliability issues inherent in task setup, such as rate limits on third-party dependencies in external registries used by the verifier. More details outlining these updates and other findings are available as part of our Laguna M.1/XS.2 Technical Report
We used the highest publicly-referenced scores for all comparison models across each benchmark.