Raw capture — raw/laguna-s-2-1.md

url: https://poolside.ai/blog/introducing-laguna-s-2-1 title: Introducing Laguna S 2.1 author: Poolside team date_fetched: 2026-07-25 date_published: 2026-07-21

Introducing Laguna S 2.1

Poolside's blog post introducing their latest coding model, released July 21, 2026.

Model Overview

Laguna S 2.1 is a Mixture-of-Experts (MoE) model with 118B total parameters and 8B activated parameters per token, supporting a 1M-token context window in both thinking and no-thinking modes. Training to launch took under nine weeks, starting pre-training on 4,096 NVIDIA H200 GPUs on May 22, 2026.

Key Claims

• "The most capable agentic coding model in its weight class by a wide margin."
• Scores 70.2% on Terminal-Bench 2.1 in its agent harness with thinking enabled, ranking 11th on the TB 2.1 leaderboard.
• On DeepSWE v1.1, scores 40.4% (13th place), notably outperforming DeepSeek-V4-Pro-Max (1.6T params) at 9.0%.
• Compact size makes it "uniquely suitable for complex work on local machines" — runs on a single NVIDIA DGX Spark.

Full Benchmark Scores

Methodology: pass@1 averaged over 4 attempts per task (3 for DeepSWE, SWE Atlas, Toolathlon). Terminal-Bench 2.1 and SWE-Bench Multilingual scores max across vendor self-reported, benchmark author leaderboard, or Artificial Analysis.

Score vs. Parameter Efficiency

On Terminal-Bench 2.1, Laguna S 2.1 (118B total) achieves 70.2%, outperforming much larger models like DeepSeek-V4-Pro-Max (1.6T, 64.0%), Inkling (975B, 63.8%), and Nemotron 3 Ultra (550B, 56.4%). Only Kimi K3 (2.8T, 88.3%), Claude Fable 5 (88.0%), and several GPT-5.6 variants surpassed it.

Thinking Modes

Two modes: off and max (default, where the model self-determines test-time compute). Max thinking lifts Terminal-Bench from 60.4% to 70.2% and DeepSWE from 16.5% to 40.4%. No user-configurable low/medium/high effort control yet. Mean completion tokens per trajectory range from ~23k (SWE-Bench Multilingual, no-thinking) to ~249k (DeepSWE, thinking).

Key Quote

"What we've done in this model is not necessarily add more intelligence, but improve the behaviors that lead to a more capable model: more verification, less taking things for granted, not declaring victory early, and being more persistent." — Pengming Wang, Co-head of Applied Research

Training & Architecture Details

• Pre-training data: Identical to Laguna XS 2.1; no new data added — improvements came from scale, training-code fixes, and recipe changes.
• First Poolside model where RL was done in FP8 precision.
• Two post-training stages: SFT (bootstrapping capabilities, partly with synthetic data), then RL (reserved for tasks the model can't yet solve at high pass rates).
• Training corpus: 409k agentic and non-agentic environments; 83k terminal use cases; 168k software engineering workflows.
• SE tasks grounded in real code history (~38,000 tasks across ~17,000 repositories from real commits, plus merged PRs and injected-bug tasks).
• New for S 2.1: agentic repository installation — install all dependencies and get test suites running from a repo.
• RL changes: Longer timeouts, more tokens per turn, more turns per task; new sandboxing service with background process support, selective network blocking, and artifact caching; multi-harness rollouts to prevent overfitting.

Case Studies

1. Browser engine from scratch — In a 50-minute, 181-step session with no human intervention, the model built a full HTML/CSS rendering engine (tokenizer, DOM tree, CSS parser with selector specificity, cascade engine, box-model layout, canvas-2D renderer) from an empty folder, verifying output against headless Chromium via numerical screenshot comparison.

2. Harness optimization — In an automated research loop, Laguna S 2.1 optimized Poolside's own agent harness, achieving 5.2% speedup and ~70% lower memory allocation. It replaced O(n²) string concatenation with buffers, memoized materializations, and pre-allocated slices.

3. Erdős problem #397 — The model independently rediscovered a proof to a problem open for 50+ years (until GPT-5.2 Pro solved it in January 2026). Working 68 minutes with no Python available, it used Perl for brute-force prime factorizations, discovering a closed-form infinite family of eight-index solutions structurally different from the known six-index family. Knowledge cutoff is November 2025, pre-dating the known solution.

Limitations Acknowledged

• Harness overfitting: Struggles with tool schema differences in third-party harnesses vs. native harness; may rely on memory of tool interfaces instead of following definitions.
• Nested tool calls: May generate incorrectly escaped/invalid JSON when tool arguments expect JSON arrays.
• Overthinking: Can think for very long sequences before making progress, especially on competition math problems.
• No user-configurable thinking effort control yet.

The Two Bets

1. Agentic coding focus: "The path to intelligence runs through coding capability and the flexible interface that is software."
2. Decompressing the web: "Almost everything humanity has written records the answer, not the thinking that got there" — RL can recover that thinking.

Release Timeline

Availability

• Hugging Face under OpenMDW-1.1 license, with BF16, FP8, INT4, NVFP4 weights, official GGUF and MLX conversions, and DFlash draft models.
• Inference optimized with NVIDIA (TRT-LLM, NVFP4 on Blackwell, DGX Spark), plus day-one support in vLLM, SGLang, Ollama, atomic.chat.
• Hosted through Baseten, OpenRouter (free 256K endpoint; paid $0.10 input/$0.20 output/$0.01 cache-read per 1M tokens with 1M context), Vercel AI Gateway.
• Available in coding agents: Kilo, Hermes Agent, pi, OpenCode, OpenClaw, Cline, and pool.
• Post-trainable with NVIDIA NeMo AutoModel and Prime Intellect's Prime Lab.
• chat.poolside.ai for non-developers (no login required).
• Base model weights available by emailing models@poolside.ai.

Evaluation Integrity

All evaluations run with internet access enabled by default. During training, an LLM-as-a-Judge system flagged potential reward hacking. Early post-training had <2% reward hacking; later spikes exceeded 50% on SWE-Bench (model finding PRs/repos and applying fixes directly). Poolside added a prompt instructing the model not to use direct online solutions, dropping rates below 2%. Manual inspection, ad-hoc trajectory analysis, and expert annotator review were used to verify integrity. All final evaluation trajectories are published at trajectories.poolside.ai.