← Back to feed

Poolside has released its first public models: Laguna M.1 and Laguna XS.2, both mixture-of-experts architectures trained from scratch on 30T tokens and optimised specifically for agentic coding workflows.

The company spent two years operating under the radar, serving government and public-sector clients who required on-prem and air-gapped deployments. The Laguna release is Poolside’s first move into the open market.

Benchmark Numbers

Laguna M.1 (225B total, 23B activated):

• SWE-bench Verified: 72.5%
• SWE-bench Multilingual: 67.3%
• SWE-bench Pro: 46.9%
• Terminal-Bench 2.0: 40.7%

Laguna XS.2 (33B total, 3B activated):

• SWE-bench Verified: 68.2%
• SWE-bench Multilingual: 62.4%
• SWE-bench Pro: 44.5%
• Terminal-Bench 2.0: 30.1%

For context: the SWE-bench Verified frontier sits at 88.7% (GPT-5.5) and 87.6% (Claude Opus 4.7). Laguna M.1’s 72.5% puts it below the current frontier tier but roughly level with Qwen3.6 35B-A3B (73.4%). Where Poolside differentiates is the SWE-bench Pro score — 46.9% on M.1 is competitive with mid-tier frontier models on harder, real-world coding tasks.

Architecture

Both models use MoE routing with sigmoid gating. XS.2 adds a hybrid sliding window / global attention layout in a 3:1 ratio across 40 layers, with a 512-token local attention window and FP8 KV cache quantisation to reduce memory per token. Context window: 131,072 tokens. Native reasoning support is built in — interleaved thinking between tool calls, per-request togglable.

M.1 was trained on 6,144 interconnected NVIDIA H200 GPUs and completed pre-training at the end of 2025. It serves as the base for the full Laguna family.

What Poolside Pushed On

Three technical bets defined the Laguna training:

Muon optimizer. Poolside replaced AdamW with a distributed Muon implementation across all training stages. Result: same training loss in ~15% fewer steps. Muon also requires one optimizer state per parameter versus two for AdamW, reducing checkpoint memory. During M.1 pre-training, total optimizer overhead was under 1% of step time.

AutoMixer. Instead of hand-tuning data mix ratios, Poolside trained ~60 proxy models on different data compositions and fit regressors to predict downstream performance from mix proportions. The learned map drives automated optimisation. Synthetic data makes up ~13% of XS.2’s final training mix, from a total of 4.4T+ synthetic tokens across the Laguna family.

Async on-policy agent RL. Actor processes pull tasks, spin sandboxed containers, run the agent against each task, score trajectories, and write results to Iceberg tables — while the trainer runs in parallel on the most recent records. Weight transfers between training and inference nodes run over GPUDirect RDMA, moving hundreds of gigabytes of BF16 weights in under five seconds. Off-policy training stability is handled by a CISPO variant.

Availability

Both models are live on OpenRouter (free for a limited time) and via Poolside’s own API. Laguna XS.2 weights are on HuggingFace under Apache 2.0 and supported in NVIDIA TensorRT-LLM on day one, including an NVFP4 version for Blackwell. XS.2 runs locally via Ollama on 36 GB of RAM. M.1 weights are available to researchers and institutions on request at models@poolside.ai.

Poolside is also releasing “pool” — a lightweight terminal agent and ACP client/server that serves as the harness used internally for RL training and evaluation.

What It Means

XS.2 is the more significant story for developers. A 33B MoE model with 3B activated parameters, Apache 2.0 licensed, that hits 68.2% SWE-bench Verified and runs on a Mac, is genuinely useful for local coding agent workflows. That’s a class of model the ecosystem has been short on.

M.1’s numbers are below the current frontier ceiling for closed models, which limits its appeal for pure benchmark optimisation. The commercial case is that Poolside has enterprise and government security packaging that frontier labs don’t offer. The Laguna family exists on a different deployment track than Anthropic or OpenAI — one where the data never leaves the building.